Context-Aware

Intelligence in Resource-
Constrained IoT Nodes:
Opportunities and
Challenges
Baibhab Chatterjee Shreyas Sen
Purdue University Purdue University

Ningyuan Cao and Arijit Raychowdhury

Georgia Institute of Technology

extend beyond 500 bil-

Editor’s note: lion devices by 2030 [1].
This article provides an academic perspective of the problem, starting with
a survey of recent advances in intelligent sensing, computation, commu- Background and
nication, and energy management for resource-constrained IoT sensor motivation
nodes and leading to a future outlook and needs. The dynamic nature
—Shreyas Sen, Purdue University of the IoT devices, cou-
pled with their stringent
 The personal, healthcare, and ­consumer resource constraints (small-size and portability
electronic industries have experienced rapid requirements, leading to smaller energy backup, less
advancements in the past few decades due to ag- computing and memory resources in harsh environ-
gressive technology scaling and low-power, low- ments, varying channel conditions, asymmetric data
cost implementation of sensor electronics built rates in uplink and downlink, etc.) and the availability
on mobile computing/communication platforms of multiple communication modalities [wired, prox-
having small form factors. This has resulted in a imity, low-­energy Bluetooth (BTLE), ANT, LoRa, Zig-
pervasive growth of connected devices, leading Bee, Human Body Communication (HBC), MedRa-
to what is known today as the Internet of Things dio, and millimeter-­wave (mm-wave), to name a few]
(IoT), as shown in Figure 1a. Increased fidelity necessitate proper selection of communication archi-
and higher bandwidths are expected to result in tecture based on the application and corresponding
50 billion connected devices, generating 30+ ex- resource constraints. In addition, the power cost of
abytes of data per month by 2020, which would communication (≈1 nJ/bit in standard wireless net-
works [2]) may warrant intelligent allocation between
local and remote computing resources, which would
Digital Object Identifier 10.1109/MDAT.2019.2899334 require context-aware operation corresponding
Date of publication: 13 February 2019; date of current version: to different scenarios, leading to minimum energy
11 April 2019. consumption for a certain amount of infor­ mation

March/April 2019 Copublished by the IEEE CEDA, IEEE CASS, IEEE SSCS, and TTTC 2168-2356/19©2019 IEEE
7
Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 SI: Survey

cases (e.g., tactile internet, autonomous driving, and

medical emergencies). The situation demands truly
intelligent devices that are aware of the operating
conditions and contexts and can dynamically adapt
itself for optimal energy efficiency and performance
by switching among different modes (­computation-
heavy, communication-heavy, high-security, low-
power, etc.). This is shown in Figure 1b in the form
of our vision for a secure, context-aware, adaptive,
resource-­constrained yet intelligent IoT device, repre-
sented as a combination of multiple sensing, compu-
tation, and communication modalities with different
power and performance. Parts of the context infor-
mation can be generated in the cloud (for latency-
relaxed applications), whereas the latency-limited
context assessment needs to happen in the sensor
node itself, using smart learning algorithms.

Challenges in asymmetric IoT networks

Before delving into the implementation details of
contextual, adaptive machine intelligence, let us dis-
Figure 1. (a) IoT at the juncture of Moore’s cuss the specific challenges for a generic scenario of an
law (more computation enabled by IoT ecosystem that contains a multitude of heterogene-
technology scaling) and Shannon’s law ous connected devices (Figure 2). These IoT devices
(more data rate enabled by modern wireless include resource-constrained and resource-rich nodes,
standards), consuming a prohibitively high gateways, and cloud data centers. The focus of this arti-
amount of energy [2]. (b) Our vision for an cle is on the resource-constrained leaf nodes that are
RC-IoT device for optimum energy efficiency. defined in [4] as the ones that do not have the hard-
ware and software capabilities to support the Transmis-
transfer, thereby improving the lifespan of the sion Control Protocol (TCP)/IP protocol suite.
network. Also, energy-resolution scalable sensing 1) Finite Resources: In view of the IoT ecosystem,
technology [preferably in the compressed domain a resource can either be physical (such as m ­ emory,
(CD)] enables consumption of sensing energy when computation power, energy, and n ­etwork band-
required and present-day context-­ agnostic IoT sys- width) or virtual (software procedures to ­perform
tems are typically overdesigned to take care of all data compression, outlier detection, etc.). An IoT de-
possible contexts/scenarios, which trade off ­fidelity vice may lack one or more of these aspects ­because
with power consumption and, hence, degrade of size limitation and specific a­ pplications.
the energy efficiency. Cloud computing is usually 2) Heterogeneity: An IoT subsystem, such as a
employed in a larger system that enables data ana- smart home, can have a significant amount of het-
lytics, remote device monitoring, visualization, and erogeneity with respect to hardware and software
client delivery [3]. This requires the IoT nodes to [4]. Various degrees of resource constraints may co-
upload the digitized data from the sensors to the exist in the same ecosystem, which makes context-­
gateway/cloud. The cloud then ­performs data ana- awareness a challenging task since it now becomes
lytics and notifies specific management systems a function of application, device location, available
(energy, memory, and real-time OS) to take suitable computation/memory resources, channel condi-
actions. However, implementing the entire compu- tions, and communication modalities. For exam-
tation framework in the cloud would mean a higher ple, smartphones with relatively high computation
communication payload at the edge device, which in power can support advanced learning and data
turn leads to higher communication power. Also, the compression algorithms, while a small temperature
closed loop (from the sensor to cloud and back to the sensor must resort to elementary learning and data
sensor) latency might be prohibitively large for certain processing methods [5].

8 IEEE Design&Test

Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 3) Security: IoT systems envision automatic dis-
covery and support for a new target device without
human intervention [6], [7], which immediately
raises concerns on the security and privacy aspects of
the network. With the limited resources and l­atency
constraints, proper authentication and/or authoriza-
tion mechanisms become all the more challenging,
as the devices with fewer resources often tend to sac-
rifice security for lower energy consumption.
4) Context and context-based adaptability: To
optimize the performance of the individual nodes,
specific contexts/modes need to be defined as dis-
cussed in the previous section, with a proper switch- Figure 2. IoT ecosystem and its specific
ing arrangement between modes for adaptability. challenges [4].
The definition of context, as will be discussed later,
is highly application dependent, and therefore the International Telecommunication Union, IoT is
implementation of context-based adaptability would a vision that ensures “from anytime, anyplace
be different for every application and either needs to ­connectivity for anyone—we will now have connec-
be decided beforehand by the designer or learned tivity for anything.”
on the fly by the system. 2) Machine intelligence: Machine intelligence
5) Scalability and reconfigurability: In addition, is usually associated with Machine learning (ML),
the IoT ecosystem should be capable of handling a which is defined in [11] as “the adoption of com­
variable number of nodes due to the mobility and putational methods for improving machine perfor-
dynamic properties of the devices, and the hardware mance by detecting and describing consistencies
and software implementations should be scalable and patterns in training data.” In view of the resource-­
to a large population of devices. It is important to constrained IoT (RC-IoT) nodes, however, intelligence
note that the previously described challenges also or edge intelligence refers to the process of context
create asymmetry among the nodes in the network, discovery and assessment, which is imperative in the
as there could be a need for communication b ­ etween realization of context-aware, adaptive techniques and
two devices with unequal resources and capabili- strategies (hardware/algorithmic/learning-based) for
ties. Indeed, IoT has a communication bottleneck in sensing, computing, and communication in the con-
the uplink, as typical IoT applications (smart sens- strained environment.
ing, wearable devices, healthcare, etc.) involve up- 3) Resource: Adopting the generic, all-encom-
loading the collected data from multiple sensors to a passing definition [12], a resource is defined as “any
single base station [5]. This asymmetry can be object which can be allocated within a system.”
optimally leveraged with a high-level goal to reduce For IoT systems, the most important resources are
the energy consumption of the overall system, as will memory (for storage), energy (for battery lifetime),
be explained in the following sections. compute c ­ apability (for computation), and network
bandwidth (for communication). Depending on the
Common terminologies used throughout available memory, RC-IoT devices are categorized
the article into Class-0, Class-1, and Class-2 devices as shown
1) IoT: Small-scale developments of internet- in Table 1 [13], with Class-0 devices having the most
connected devices were materialized as early as stringent constraints.
1982, when researchers at Carnegie Mellon Universi- 4) Context and context awareness: The notion of
ty deployed a Coke vending machine with an online context-aware computing was first introduced by Schilit
inventory [8]. Mark Weiser’s famous 1991 paper on and Theimer [14]. Although many definitions exist for
ubiquitous computing [9] envisioned the concept of context and context awareness, the one provided by
a large scale implementation, and the term Internet Abowd et al. [15] is widely accepted as a concrete
of Things was coined by Ashton in a presentation at definition of context based on the five Ws (who,
Proctor and Gamble in 1999 [10]. According to the what, where, when, and why). As has been argued

March/April 2019
9
Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 SI: Survey

5) Quality of context: Quality of context (QoC) is

Table 1. Available resources for RC-IoT Devices [13]. related to how well the context model and its a
­ ttributes
are extracted from raw sensor data. QoC is defined
using a combination of parameters such as validity,
precision, and update rate of the context information,
which is ­processed out of the raw data from a sensor.
Reference [6] presents a detailed ­survey on QoC.
in [16], all previous definitions [17]–[20] of context Building toward the concept of a context-aware,
and context-awareness suffered from the specificity adaptive RC-IoT system, we shall discuss intelligent
of the example applications that they were referred hardware techniques for sensing (compressive sens-
to and could not be used to define the new context. ing, time-based sensing), computation (edge ana-
References [15] and [16] defined ­context as follows: lytics/in-sensor analytics in the form of anomaly
Context is any information that can be used to detection and data compression), communication
characterize the situation of an entity. An entity [Intraphysical layer (Intra-PHY) and Inter-PHY adap-
is a person, place, or object that is considered tation, along with two recent sub-10pJ/b communica-
relevant to the interaction between a user and tion modalities, i.e., proximity communication and
an application, including the user and applica- HBC], and energy management (dynamically recon-
tions themselves. figurable LDO, switched-mode LDO, and intermittent
powering) in this article. We shall also present two
In light of the above definition of context, context- examples of cross-layer adaptive systems that employ
awareness is defined as follows: “A system is context- more than one approach discussed in this article for
aware if it uses context to provide relevant information optimum power-efficiency. Finally, after describing
and/or services to the user, where ­relevancy depends on various security considerations and learning tech-
the user’s task”[15].A context is usually represented by a niques for RC-IoT devices, we present our view of the
model and its attributes, and it is often described based current state-of-the-art and where it needs to be in
on the application scenario. Further details on this can the near future, based on the learnings and research
be found in [6] and [21]. in the domain of secure, context-aware, adaptive
Rc-IoT nodes over the last two decades.

Intelligent sensing
Compressed-domain signal acquisition
Compressed-domain sensing/compressive sens-
ing (CS) [22], [23] is a mathematical tool in sig-
nal processing that defies the Shannon–Nyquist
sampling theorem by sampling a sparse signal at
a rate lower than the Nyquist paradigm and still
being able to reconstruct the signal with negligi-
ble errorrate (Figure 3). Since its inception, CS
has found m ­ ultiple applications including image
processing [24], m ­ edical imaging [25], RADAR
technology [26], in-sensor analytics [27], gesture
recognition [28], [29], and healthcare [30]. CS
algorithms assume that the signal to be sampled
has a sparse representation, and it was shown that
sparse signals with randomly [from independent
and identically distributed (i.i.d) Gaussian dis-
tribution] undersampled data can be recovered
with a low error by formulating it as an optimi-
Figure 3. Nyquist rate sampling/sensing versus zation problem. Hence, the advantage of CS is
compressive sensing. twofold: 1) CS allows a lower sampling rate that

10 IEEE Design&Test

Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
reduces the power consumption in the analog-to-
digital converter (ADC) and clock generation
circuitry, and 2) compression creates a smaller
amount of data with rich information-­content that
reduces the burden on the subsequent processing
and communication modules. Since many of the
naturally occurring signals such as sound, visual
image, or seismic data can be represented in the
sparse form [31], it is possible to leverage the supe-
rior energy efficiency of CS in an IoT scenario. Two
comprehensive reviews on CS can be found in [31]
and [32].
To mathematically represent CS more clearly,
let us assume that the orthogonal basis {​​​​ψ​  i}​​ ​​  n ​​​ span
i=1
the n-dimensional real space ℝn. Then, any signal​
x ∈ ​R​​  n​​can be represented by matrix multiplication
of the matrix ψ with the elements of a sparse vector
n
S = [S1, S2, S3, ..., Sn]T ∈ Rn such that ​x = ​∑ i=1​​  ψ​  i​​​ ​Si​  ​​​. If
the vector S has only k ≪ n nonzero entries, then the
signal x is said to be k-sparse, and ψ is called the spar- Figure 4. CS: creation of an m × 1 meas-
sifying/representation matrix for x. For CS, the n × 1 urement vector from an n × 1 signal of
input signal x is pre-multiplied by an m × n sensing interest (m < n).
matrix Φ to get an m × 1 compressed signal y, where
m < n and the ratio (n /m) is termed as the com- Amaravati et al. [33] exploit CD data processing,
pression factor. This is represented by the following which allows trigger detection with significantly
equation and is shown in Figure 4 lower power and computational requirements. This

y = Φx = Φ ψ S.​
​ (1)
__
If the coherence (correlation) ​ μ​(Φ,ψ)​ = ​√n ​  max
​|​Φ​  j​​ , ​ψ​ i​​|​​ (where 1 ≤ i ≤ n and 1 ≤ j ≤ m with Φ​ 
​​ j​​​ being
the jth row of Φ) is low, it can be proved that fewer
samples are required to reconstruct the signal [23].

Compressed domain processing and

computational data converters
Multiple CS algorithms for IoT applications
have been developed in the last decade. Reference
[34] showed a matrix-multiplying ADC (MM-ADC)
in 130-nm CMOS technology and demonstrated two
applications: 1) electrocardiogram (ECG)-based car-
diac ­arrhythmia detection (9.7× energy savings as
compared to ­traditional ADC followed by arrhyth-
mia detection) and 2) image-pixel-based gender
­detection (23× energy savings as compared to tradi-
tional ADC f­ollowed by gender detection). Feature
extraction and ­ classification were combined in a
single measurement matrix (Φ in CS theory) for light-
weight applications as shown in the work. Our ear-
lier work [33] demonstrated a light-powered smart
camera with CD gesture detection. To enable always Figure 5. Block diagram of the CS algorithm
on and self-powered operation on IoT devices, for gesture detection [33].

March/April 2019
11
Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 SI: Survey

In this system, the gesture motion is captured by

a sequence of difference images between consecu-
tive frames. Each difference image passes through
two layers of compression to reduce its resolution
and to be transferred to the CD. In the first layer, the
resolution is reduced by dividing the whole image
into several blocks and taking the average of each
block. In the second layer, coded combinations of
these block-averaged pixels are extracted. It then
estimates the center of the motion directly from
these compressed measurements. These motion
centers are passed to a classifier for gesture recog-
Figure 6. Accuracy of the detection of
nition. Figure 5 shows the block diagram of the pro-
hand gestures “Z,” “X,” “O,” “N,” and “+”
posed system, while Figure 6 presents the accuracy
as a function of the irradiance level of the
of detecting different hand gestures as a function of
light-powered smart camera with CS [33].
the irradiance levels in the environment where the
camera operates. In the work presented in [33], the
is in contrast to existing algorithms that work directly sparse compression algorithm was performed in
in the pixel domain. Given the objective of the cam- the microcontroller unit instead of the ADC. In [27],
era front end (FE), the computation complexity can the authors have shown an ASIC implementation in
be largely reduced (768×, as demonstrated in [33]) 130-nm technology that utilizes CS DAC and MM-ADC
from existing algorithms that are targeted for continu- together to achieve only 165-nJ/frame classification.
ous gesture recognition [35], [36]. Figure 7 shows the results from an arrhythia detec-
tion ASIC [30] with a time-based CS ADC. A total of
160 parallel processing units were employed on-chip,
and an accuracy of 84% was achieved with only
10.5-nJ energy per classification for a compression
ratio of 8×. One key idea here is the introduction of
computational ADCs, where analog input signals are
not only digitized but also computed upon during
acquisition. In particular, computational ADCs pro-
vide linear transformations of the signal in a single
stage, thus improving the system energy-efficiency.
In a more recent work [37], a submicrowatt CS
hardware is presented in 65-nm CMOS technology
with online self-adaptivity for incoming signals with
varying sparsity. Initial efforts of self-adaptivity were
earlier demonstrated in [38] using an asynchronous
ADC with an adjustable sampling rate and in [39]
using temporal decimation and wavelet shrinkage.
Both of these techniques were utilized with specific
incoming signals. On the other hand, Roose et al. [37]
offer a more general technique that exploits the online
Figure 7. (a) Arrhythmia detection using time- sensory data statistics for dynamic reconfiguration
based CS ADC with embedded classification and (in terms of the compression algorithm, compression
INL-aware training [30]. (b) Classification accuracy harshness, and sampling frequency).
versus compression ratio. (c) Energy efficiency
(nJ/classification) versus compression ratio, show- Sensing using time/frequency
ing that a compression ratio of 8× achieves ≈90% Many of the naturally occurring signals in IoT
lower energy with an accuracy hit of <5%. are slowly varying, such as temperature, humidity,

12 IEEE Design&Test

Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
and vibration. Most of the energy content in these
signals is contained within extremely low frequen-
cies. The resolution and dynamic range (DR)
requirements for these applications, however, can
be small (e.g, temperature and humidity), large (e.g,
vibration), or variable based on environment (e.g,
radiation). Voltage-mode and current-mode ADC
­
designs in these scenarios become limited by the
ambient noise, supply rails, and power consump-
tion. Time-based ADCs, on the other hand, can uti-
lize the availability of time (since signals are of very
low frequency) in an energy-resolution scalable
manner as shown in [40]. For high-resolution require-
ments, the signal to be sensed is converted into an
equivalent frequency (using a resistive sensor and
a ring oscillator-based resistance to frequency con-
verter) and is simply observed (using a counter) for
a longer amount of time for a change in the aver-
age frequency. For low-resolution requirements, the Figure 8. Working principle of time-based ADC:
frequency is observed for a shorter amount of and higher resolution with more integration time and
then can be turned off (through duty cycling) for frequency/energy-resolution scalability as
saving energy. Figure 8 shows the working principle compared to traditional ADCs [40].
for the time-based ADC for detecting the difference
between a frequency f and its slightly modified ver-
sion f 2. The minimum amount of time for which we
need to observe/count the frequencies to detect the
difference is ​​1 ⁄ |​​f​  1​​  − ​f​  2|​​ ​ ​​. Hence, for a smaller |​​​f​  1​​  − ​f​  2​​|​​
(high-resolution requirement), the time to enable
the counter needs to be higher.
Even though this method ensures energy-­
resolution scalability within a range, the resolution
cannot be made infinitely high by waiting for a
longer time. The ambient noise statistics, process,
voltage, and temperature (PVT) variation, and jitter
accumulation in the ring oscillator would limit the
achievable resolution, out of which jitter accumu-
lation is shown to be the dominant factor in [40]
in a controlled environment for radiation measure-
ment. This is demonstrated in Figure 9, where it is
shown that the scaled quantization error in measur-
ing a fixed frequency within a p ­ redefined amount
of time goes down with the time of measurement.
­However, the accumulated jitter from the ring oscil-
lator goes up with the total time of measurement.
If the slope of the linear plot of accumulated jitter Figure 9. Application of time-based ADC in radiation
versus m­ easurement time is k, then the achievable sensing [40] using a resistive floating gate sensor, a
resolution is shown to be limited to log2 (1/k) bits. three-stage differential ring oscillator, and counters.
The system in [40] achieves 18-bit resolution 18-bit resolution is achieved with 861-nW power,
with 861-nW power consumption (one reading per utilizing the tradeoffs among measurement time, bit
second) and 12-bit resolution with 9.04-nW power resolution, and accumulated jitter.

March/April 2019
13
Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 SI: Survey

consumption (one reading per second). The resolution analysis. However, even if a fictitious technology
can be improved by phase noise reduction techniques could potentially offer zero capacitances, a zero-
for the ring oscillator at the cost of a higher power. power receiver (Rx), and 100% efficiency for the trans-
mitter (Tx), (Ecomm/bit) would still be limited by the
Collaborative sensing free-space path loss (PLFS) of the physical channel,
Collaborative wireless sensor networks [41]–[43] which is given by Frii’s equation [48], [49] and shown
can sense an analog signal over a large-area test bed in the following:

P ​LF​  S​​ = ​GT​  x​​  ⋅​GR​  x​​  ( ​  λ  ​)m

(e.g., soil nitrate sensing for smart agricultural appli-
​ ​ ___ ​ ​, (4)
cation) utilizing collaborative efforts among the sen- 4πd
sor nodes and their communication with each other/ where GTx and GRx represent the gains of the trans-
with the cloud, which brings us to the next part of mitting and receiving antennas, respectively, λ is the
the article—the tradeoffs and power optimization wavelength, d is the distance between the Tx and Rx,
among in-sensor c ­ omputation, short-range commu- and m is a parameter (typically between two and
nication, and long-range ­communication. three) that represents the fading margin. For d = 10 m
and a typical ANT protocol operating at 915 MHz, the
Intelligent computing platforms most optimistic PLFs (m = 2, GTx = 2 dB, GRx = 2 dB)
As the number of distributed sensors and IoT turns out to be about 48 dB, which means, with a state-
end-nodes are increasing, the total amount of data of-the-art Rx sensitivity of −100 dBm [50], the Tx needs
transfer to the backend cloud servers are becoming to transmit a minimum of −52 dBm. This translates to
prohibitively large, resulting in network congestion a power consumption of 6.3 nW (theoretical limit—
and high energy consumption during data trans- assuming no capacitance and 100% efficiency) and
mission at the sensor node [44]. This motivates the an energy/bit of 105 fJ/b (for a ­maximum data rate of
need for in-sensor data analytics that would perform 60 kbps for ANT), which is 107 times higher than the
context-aware data acquisition with compression, theoretical minimum energy/bit for computation, as
followed by transmit if necessary. given by the Landauer principle.
From the foregoing analysis, the theoretical min-
Need for intelligent computing imum energy/bit for communication is given by the
The computation and communication energies physical limits of the channel, and can be written as
(Ecomp and Ecomm, respectively) in a system can be R ​x​  ​​
​​​(​Ec​  omm​​  /bit)​​  th_min​​  = _____________________
​   
   sen
 ​​,(5)
​(​GT​  x​​ ⋅ ​GR​  x​​ ​​(___
​  λ  ​)​​​  ​)​  × η .DR
written as n
4πd
​​  E​ comp​​  = ​(​Ec​  omp​​  / bit)​ × Number of bits switched
where Rxsen = kBT50Ω × NF × SNR × k × DR is the Rx
​Ec​  omm​​  = ​(​Ec​  omm​​  / bit)​ × Number of bits sent.​ (2) sensitivity as a function of DR [51], η is Tx efficiency,
and DR is the data rate supported.
(Ecomp/bit) will be dominated by the dynamic
Figure 10 shows the comparison of Ecomm and
power for frequencies above the leakage-dom-
Ecomp for the same number of bits transmitted, or
inant region, as shown in [45] and [46], and,
switched. The state-of-the-art wireless transceivers
hence, can be approximated by (Ecomp/bit) = CV 2,
[52] consume ≈104 times more energy as compared
which scales with technology. If an ideal technol-
to computational bit switching in 45- and 65-nm
ogy had allowed zero device capacitances, then
nodes. This bottleneck analysis directly signifies that
(Ecomp/bit) would be very close to the theoreti-
some amount of intelligent computation at the sen-
cal limit posed by the Landauer principle [47] as
sor node (in-sensor analytics) would help in bring-
given in the following:
ing down the total energy by enabling selective data
​​​(​Ec​  omp​​  / bit)​​  ​​  = ​k​ B​​  T × ln 2,​ (3) transmission, which will reduce Ecomm at the cost of
th_min
additional Ecomp.
where kB is Boltzmann’s constant and T is the
ambient temperature. For room temperatures, In-sensor analytics as a form of edge
(Ecomp/bit)th _ min is calculated to be about 2.9 × 10 −21 J. intelligence
For a standard 45-nm CMOS technology node, the bit Based on the communication and computa-
switching energy was simulated to be ≈1fJ for this tion energy tradeoffs and the amount of resources

14 IEEE Design&Test

Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
available at the RC-IoT device, partial or complete
processing of the sensor data (e.g., anomaly detection
and data compression for sensor readout, and object
localization and segmentation for video surveillance)
can take place in the leaf node itself. In this section,
we discuss the two most common ISA techniques for
RC-IoT devices, namely, anomaly/outlier detection
and data compression. The anomaly detection meth-
ods can enable selective (and immediate) data trans-
mission when an anomaly occurs in an otherwise nor-
mal sensor readout. As a healthcare example, selective
ECG data t­ransmission with arrhythmia (anomaly) Figure 10. Comparison of communication and
detection would ensure immediate notification with computation energies (both theoretical and from
minimum communication cost. Data compression, standard implementations [52]) that show that
communication energy is ≈104 times more than
on the other hand, would ensure that the maximum
computation energy (with same number of bits).
amount of information between transmissions can be
Leakage power is ignored in the analysis.
stored in a small amount of on-sensor memory.
1) Anomaly/Outlier Detection: According to
the RC-IoT device itself for optimum computation-
­Barnett and Lewis [69], “an outlier is an observation
communication tradeoff.
(or subset of observations) which appears to be in-
2) Data Compression: As shown in the “Intelligent
consistent with the remainder of that set of data.”
Computing Platforms” section, compressive-sensing
Figure 11 shows an example of anomaly in a sensor
techniques can result in significant energy savings
readout and explains the three classes of anomaly
in the ADC, on-sensor processor, and communica-
that are common in IoT devices and wireless sen-
tion modules. It must be noted that CS-ADC is still
sor networks [70]. It is to be noted that the p ­ rimary an emerging technology and has not yet become an
difference between outlier detection and event
­ integral part of commercially available embedded
­detection is the fact that an outlier is detected by ­frameworks. In-sensor data compression techniques
comparing the readings from the sensor with each on the IoT processor, however, have also shown
other and without any prior semantics that define the ­energy benefits by bringing down the communica-
trigger conditions of an anomaly. On the other hand, tion power. Some of the earliest reported works on the
trigger conditions for event detection are usually tradeoff between the raw data communication and
­defined a priori, and the sensor readouts are com- the compressed data communication are from MIT’s
pared with that trigger condition to detect an event. Computer Science and Artificial Intelligence Labora-
Outlier detection algorithms utilize spatio-­ tory [72] and from CMU’s Odyssey Project [73]. Ref-
temporal correlations among the data points from the erence [73] used application-aware adaptation that
same node and/or neighboring nodes to distinguish trades off data quality with resource ­consumption
between normal operation and anomalies [70].
Table 2 shows some of the most common anomaly
detection techniques for wireless sensor networks
and IoT. These methods include both learning
(supervised and unsupervised)-based techniques
and algorithmic (statistics-based) techniques and
offer various orders of resource requirements and
accuracy. Some of the most recent works include a
hybrid statistical method from Twitter [71], which
has low latency and high accuracy but needs more
computational resources. Simplistic techniques such
as mean- and average-based statistical analysis [53], Figure 11. Example of anomaly in sensor
on the other hand, can be implemented easily on readout.

March/April 2019
15
Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 SI: Survey

Table 2. Outlier detection techniques for wireless sensor networks/IoT.

Technique Example Salient features Drawbacks

Statistics-based 1) Gaussian Parametric 1) Spatiotemporal correlation for Gaussian Simplistic and can suffer
Estimation [53], 2) Non-Gaussian nonanomalous data, fixed thresholds from low accuracy
Parametric Estimation [54], 3) Kernel for anomaly detection, 2) anomalies
Density Estimation (non-­parametric) are treated as SaS-distributed impulsive
[55], [56], and 4) histogram-based events, 3) no a priori PDF is assumed.
method (nonparametric) [57] Kernel density functions approximate
the PDF, and 4) works on histograms and
not on raw data (inherent compression—
reduced communication cost)

Nearest-Neighbor-based Euclidean distance [58] and dynam- Simple implementations for both Resource-extensive for
ic time warping methods [59] ­univariate and multivariate data multivariate data

Clustering-based Creates clusters based on raw data Can be employed to take care of incre- Resource-extensive for
and detects outliers that do not fall mental processing ­multivariate data, suffer
into any cluster [60] from the choice of an
­appropriate cluster width

Classification-based 1) SVM approach [61], [62], 1) Maximally separated classes (one/two Computationally
2) Bayesian Network approach [63], class approach to reduce complexity, 2) intensive
[64], 3) long short term memory uses Bayesian Intuitions to predict anom-
(LSTM) [65]/hierarchical temporal alies, and 3) uses LSTM/HTM for time-­
memory (HTM) [66] approach series data pattern of unknown length

Spectral-decomposition- PCA-based approach Added advantage of dimensionality Selecting suitable

based [67], [68] ­r eduction/data compression principal components is
­computation-heavy

with the help of an embedded OS. Reference [72] IoT devices include 1) principal component analy-
experimentally showed that the ratio of energy re- sis (PCA) ([75]–[77], which use lightweight PCA for
quired to transmit 1 bit is ≈480–1270 times higher than dimensionality reduction and data compression),
that of a 32-bit addition under varying channel condi- 2) coding by ­ordering ([78], where the data from one
tions. This means that a compression algorithm that node is shown to be encoded by the order at which
is able to remove more than 1 bit from a string of data other nodes in the same hierarchy communicate
would have energy benefits if the algorithm is equiv- with the parent node), 3) burst mode/pipelined tech-
alent to (or less than) 480 addition instructions. The niques ([79], where data are stored, packetized, and
standard compression algorithms explored in [72] transmitted in burst mode to remove redundancies
[such as bzip2/Burrows–Wheeler transform (BWT), and number of transmitter switch on/off), 4) frame
Lempel–Ziv–Welch (LZW), Lempel–Ziv–Oberhumer difference-based compression ([33] and [80] that
(LZO), and prediction by partial matching (PPMd)] store differences in consecutive frames for video
are much smaller than 480 additions, which means compression), and 5) distributed data compression
that any of these algorithms would be beneficial. [81] using conditional entropy encoding with corre-
However, the key limitation in an IoT implementa- lated data between two nodes that perform spatial
tion comes from the runtime memory requirement data compression through short-range communi-
for these algorithms, which is in tens of kilobytes for cation between the sensor nodes. For optimum re-
LZO to hundreds of kilobytes for BWT. This readily source utilization, this short-range communication
makes these algorithms infeasible for C0, C1, and can be a low-power ­communication scheme, such
C2 RC-IoT devices (referring to Table 1). More light- as ­MedRadio or HBC (for body area networks within
weight compression techniques, such as miniLZO a few meters), which consumes hundreds of micro-
and sensor LZW with mini cache (S-LZW-MC) [74], watts, or ANT/BTLE, which consumes a few milliwatts
require only 8.192 and 3.250 Kbytes memory, respec- to ≈10 mW when on, as will be shown in the next sec-
tively, and can be used in C2 and some C1 devices. tion. After spatial compression is done, we envision
Other important techniques for data compression in that the node with the highest amount of battery life

16 IEEE Design&Test

Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 Table 3. Comparison of state-of-the-art wireless techniques for IoT nodes [82].

Proximity HBC [84], NFC ZigBee BT/BTLE ANT WiFi LoRa WAN
comm. [83] [85]

Distance 1 mm 2–5 m 10 cm 10–100 m 10–100 m 10–20 m 30–50 m ≈1 km

through
human
body

Data rate 8–32 G 10’s of M 20–400 k 20–200 k 0.8–2.1 M 60 k 300 M (802.11 g), 200 k
(bps) 7 G (802.11ac/11d)

Energy effi- 4p 6.3 p 1n 5n 15 n 10 n 5n 1µ

ciency (J/b)
Security High High Medium Low Low Medium Medium/High Low/Medium

(or the node that is closest to the Rx) would take the [2]. The concept of Intra-PHY and Inter-PHY communi-
responsibility of sending the compressed data to larg- cation is presented in Figure 12a, where the switching
er distances, possibly through a high-power commu- of PHY is shown to occur based on communication
nication protocol such as LoRa WAN. distance (as an example of context), while the adap-
tation within a PHY is performed for optimum energy
Intelligent communication efficiency based on the operating conditions.
Continuous device scaling over the last few dec-
ades have resulted in cheap computation through Intra-PHY channel-adaptive radios
Moore’s law, and the ability to support higher data For Intra-PHY adaptation, the energy-­performance
bandwidths has created cheap wireless commu- tunability knobs are dynamically optimized without
nication paradigms through Shannon’s law. How- changing the PHY. Traditional techniques of scal-
ever, the progress in battery technology has been ing the energy consumption over varying channels
relatively slower, making the available energy one involve adaptive modulation and coding [87], which
of the most sought after resources in modern IoT sys- increases the order of modulation (from QPSK to
tems, thereby motivating the research needs toward 16-QAM to 64-QAM) as the channel quality becomes
low-energy sensing, computation, and communica- better and corresponding error vectors become
tion. As supported by the analysis presented in [86] more and more manageable. Although this increases
and in the “Intelligent Computing Platforms” section, the spectral efficiency of overall transmission, the
the energy cost per bit for communication is 103–104 power consumption of the radio frequency (RF) FE
times higher than the energy cost of computation effectively remains constant. As shown in [88], 70%–
for raw data bits. In the vision of the truly intelligent 90% of the overall power in a low-power transceiver
IoT nodes presented in this article, most optimum (Tx+Rx) system is consumed in the Rx FE/Tx power
energy efficiencies are expected from the commu- amplifier (PA) and LO generation subsystems, and,
nication subsystems based on the specific operating hence, significantly more energy efficiencies can be
conditions/context (such as communication dis- obtained by dynamically scaling the FE power and
tance, channel conditions, latency, quality of ser- performance according to the application. Most of
vice requirements, data rate, battery conditions, and the research efforts in building channel-adaptive
process variation) when turned on. Table 3 shows designs are concentrated toward the Tx PA and
the state-of-the-art communication modalities avail- employ techniques such as digital predistortion, Tx
able for IoT devices, which range from 4-pJ/b prox- power control, envelope tracking, polar implemen-
imity communication for ≈1-mm distance to 1-µJ/b tation, and dynamic companding with PA bias con-
long-range (LoRaWAN) communication to ≈1 km. trol [89], [90]. Rx circuit-level adaptation techniques
We readily notice the possibility of optimizing the include automatic gain control and field-program-
communication framework within a modality and mable low-noise amplifiers (LNA) with power-linear-
among different modalities, hereinafter called Intra- ity tradeoff [91]. Some of the recent advancements
PHY and Inter-PHY communication as explained in include an adaptive DR and BW Rx [92] that use a

March/April 2019
17
Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 SI: Survey

Figure 12. Vision for adaptive communication in IoT [2]. (a) Context-aware communication PHY,
which can adapt to its surroundings to perform more efficiently with experience by self-learning
the optimum operating points. Adaptation can be intra-PHY or inter-PHY based on context, indi-
cating the need for incorporating incorporating multiple adaptive PHYs per device. (b) Today’s
worst case design philosophy. Circuits/systems are generally designed to handle the worst case
conditions plus a guard band. This leads to significant loss in energy efficiency. (c) Dynamically
channel-adaptive radio (ViZOR). (d) Process-variation tolerant ViZOR (Pro-ViZOR). (e) Self-learning
energy-scalable wireless systems.

programmable gain amplifier (PGA) and an adaptive FE to dynamically optimize power and performance.
intermediate-frequency filter. ­Discrete-time spectrum If the tuning knobs are designed in an orthogonal
sensing was utilized in [93] to modify the modes of an manner (i.e., operation of one knob will modify only
Rx filter to achieve adaptive interference rejection. one specification out of linearity, gain, and NF of the
An interference-aware adaptive ADC was shown to FE), the controller was shown to achieve ≈3× better
adapt itself to a low-power mode in absence of any energy savings for best-case channel conditions [98],
blocker using a simple built-in spectrum analyzer [99] and can be optimized for either maximum data-
[94]. A channel adaptive ADC and a successive-­ rate (data-priority) or minimum energy (energy-pri-
approximation-register-based time-to-digital con- ority) for any channel [100], [101]. However, it was
verter (TDC) was shown for a 28-Gbps wireline sys- also reported that the adaptation control law, which
tem in [95]. These implementations have shown was fixed during design time, cannot work optimally
benefits in the standalone adaptive subsystems under manufacturing process variations. References
(such as the LNA, PGA, or TDC). However, it must [90] and [102] solve this problem by detecting the
be noted that, unlike a Tx where most of the power process corner of the device under consideration
is consumed in the PA, the Rx power consumption is using built-in process sensors and updating the
more distributed among different blocks; hence, the control law accordingly during postmanufacturing
entire Rx should be considered as a unit for power- tuning. This technique (i.e., Pro-ViZOR) is shown in
performance tradeoff analysis. It was shown in [96] Figure 12d and requires high design-time effort
and [97] that the best-case energy-savings in an Rx to cover the entire process-corner space for the
FE can be obtained by distributing the instantane- ­power-performance adaptation [103].
ous performance-slack optimally across different
building blocks in the Rx. A precharacterized con- Intra-PHY adaptation: Self-learning radios
trol law (defined during design) was employed to The high design-time complexity of Pro-ViZOR
achieve multidimensional adaptation of multiple Rx was significantly reduced by employing self-learning
components with virtually zero-margin (ViZOR) Rx wireless systems (Figure 12e) that gradually learn
operation. Figure 12c shows the operation of ViZOR the adaptation control law when the device is in idle
using design-time tuning knobs and sensors in the Rx condition [104]. Figure 13 shows how the learning

18 IEEE Design&Test

Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
algorithm populates the power-performance channel
space during an intermediate time instant and at
the final time instant when learning is complete.
Figure 14 presents the average power consumption
of such a self-learning ­channel-­adaptive wireless sys-
tem over multiple days to include various channel
conditions. The initial overhead is due to the need
for controlled on-line experiments during real-time
operation that gathers useful data points for learn-
ing the control law. It is shown that this system
becomes increasingly energy-efficient with experi-
ence [105]–[107]. When the power consumption
saturates with the learning, the overhead of con-
trolled experiments is removed (day 29 in Figure 14).

Communicating with ultralow-energy

(<10 pJ/b) PHYs Figure 13. Self-learning the adaptation control
Today’s wireless technology is limited by the high law: learning of channel space and power profile at
channel losses (≈ 60–80 dB in standard operating intermediate and final time instances [2].
conditions) that increase the power consumption in
the Tx to compensate for this channel loss. In addi- ­evices, which are brought in close proximity
d
tion, narrowband wireless techniques employed in (≈1 mm) to establish connectivity through capaci-
standard implementations involve frequency upcon- tive coupling. Because of the antennaless capaci-
version (Tx) and downconversion (Rx), which tive t­erminations in both the devices, the channel
increase the power overhead to enable smaller behaves like a simple capacitive divider with low
antennas and multiplexing. loss and maximally flat frequency response, thus
Due to these reasons, traditional wireless tech- enabling broadband signaling. One key challenge in
niques such as near-field communication (NFC), this mode of communication arises from the self-res-
Zigbee, BTLE, ANT, and Wi-Fi can only achieve a best- onance frequency (SRF) of the inductive vias with
case energy efficiency of ≈1 nJ/b, while low-power the distributed parasitic capacitance of the coupler
wireless body-area networks (WBANs) and M ­ edRadio plates. The SRF creates a peak in the channel fre-
implementations usually operate at hundreds of quency response, thus disturbing its otherwise flat
pJ/b for short-distance (1–5 m) communication [2].
­However, recent progress in wireline-like broadband
techniques can enable sub-10-pJ/b communication
over low-loss channels (mm-scale device proximity
communication, or meter-scale data transfer through
the human body), as shown with two examples in this
section. W
­ ireline-like techniques eliminate the need
for antennas as well as modulation, thus lowering the
power consumption dramatically. However, commu-
nicating with multiple devices would now require
time division multiplexing instead of frequency divi-
sion multiplexing, thereby increasing the latency if
proper scheduling techniques are not employed.
1) mm-scale proximity communication: Figure 15
demonstrates the modality for mm-scale multi-Gbps
proximity communication [83], [108] at ≈ 4 pJ/b. Prox-
imity communication is implemented by employing Figure 14. System behavior and power savings for
metal plates (couplers) in both ­ communicating self-learning radio.

March/April 2019
19
Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 SI: Survey

Figure 15. Emerging PHY example 1. (a) mm-scale proximity communication [83].
(b)–(c) Specific challenges (SRF of the interface and crosstalk). (d)–(e) Their solutions
[integrating dual data-rate (DDR) Rx that creates a notch at the SRF to mitigate ringing
and alternating rectangular differential couplers to mitigate crosstalk]. (f) Measurement
results showing BER of <10−12.

Figure 16. Emerging PHY example 2. (a) Broadband HBC is affected by interference from the envi-
ronment. (b) IR-HBC using time-domain signal-interference separation [84], enabled by 1) capacitive
termination (offers larger frequency range for broadband application) and 2) integrating DDR Rx for
interference rejection using signal-interference separation and duty-cycle adaptation. In comparison
with state-of-the-art HBC transceivers, broadband IR-HBC achieved 18× better energy efficiency
(6.3 pJ/b), which is ≈100× better than traditional WBAN.

20 IEEE Design&Test

Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
nature. An integrating r­eceiver (Rx) with a tunable
notch placed at the SRF can solve this problem as
illustrated in [83]. The other key challenge in this
technique is the crosstalk between parallel chan-
nels, which is solved ­using alternating rectangular
differential couplers that employ inherent passive
crosstalk cancelation. As shown in Figure 15e, the
crosstalk from 1+ and 1− to 2− is equal and opposite,
and hence, cancels each other, while the crosstalk
from 1− to 3+ and 3− are equal and therefore can-
cel each other differentially. Using these two tech- Figure 17. Vision for context-aware adaptive PHY. The
niques, Thakkar et al. [83] successfully demonstrate IoT node needs to store minimal information on near-
32-Gbps data transfer with bit error rate (BER) <10 −12 by devices (last transmit time, mode information, and
using four parallel channels up to 0.8-mm distance battery life) along with its own remaining battery life.
and 4-pJ/b energy efficiency (which is ≈100× lower In case of an event/anomaly detection, if the sensor
than contemporary mm-wave gigabits-per-second storage is full (even with data compression) or there
implementations [109], [110]). is a transmit timeout, the node would then asses the
2) Interference-Robust Human Body Communica- context and turn the corresponding transmit subsys-
tion (IR HBC): Many future healthcare [111], human– tem on. If a change in context requires change in Tx
computer interaction [112], [113], and n ­ euroscience modality, it will be taken care of by the context dis-
applications rely on the Internet of Body (IoB), to covery/assessment block, which can employ a struc-
connect wearable and implantable devices on, in, tured algorithm/learning framework.
and around the human body, which are typically in-
terconnected though WBAN, c ­ onsuming upward of 1 efficiencies, channel loss, data rates, and distance
nJ/b. Using the human body itself as a low-loss broad- support are incorporated in the same transceiver.
band communication medium [114]–[116], e ­nergy Figure 17 presents the vision for a context-aware
efficiencies [84], [117] similar to the proximity com- adaptive PHY that involves the following:
munication, or wireline input–output (IO) [118]–­[120]
achieve high physical security [85]. Capacitive ter- • assessment of the need for communication
mination along with voltage-mode signaling allows based on event/anomaly detection (in-sensor
broadband communication in which low loss and ab- analytics), memory (storage) buffer information,
sence of upconversion and downconversion give rise and channel quality information,
to the extreme energy efficiencies. The key challenge • context discovery and assessment based on
in broadband HBC comes from the antenna effect in ­battery life of current and nearby devices (helps
the human body that picks up unwanted interferenc- to understand which device has the most
es that corrupt the signal. An interference detection resources for long-range high-power communica-
and rejection loop using an adaptive notch [121], tion, if required),
[122] at the integrating Rx has enabled the lowest en- • last transmit time and modality of current and
ergy (6.3 pJ/b for 30-Mbps data transfer through the nearby devices (helps to understand the spatial
body, which is ≈100× lower than traditional WBAN), statistics of the data and the sensors), Rx distance
as well as the most interference robust (can tolerate and location (e.g., whether both the Tx and Rx
−30-dB signal-to-interference ratio) HBC transceiver devices are on the human body),
built to date [84], as shown in Figure 16. • the possiblity of spatial data compression based
on the information from nearby devices (if suc-
Inter-PHY adaptation: Communication with cessful, this will require long-range data commu-
context switching nication for only one node among a cluster of
Like humans, a truly intelligent RC-IoT node sensors), along with any other information from
needs knowledge (context-awareness) and adapta- the cloud.
tion according to the situation (reconfigurability).
Inter-PHY context-aware adaptation is most effective Equipped with all the knowledge, the RC-IoT device
when multiple PHYs with different orders of energy can now adapt itself to the context and transmit

March/April 2019
21
Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 SI: Survey

constant voltages, the analog and digital implementa-

tions of which are shown in Figure 18.
The power conversion efficiency (η) of an LDO
is given by
​  O​  UT,LDO​​
V ​I​  ​​
η = ​ ______ ​   × ​ ________
​ Load
 ​​ ,(6)
​VI​  N,LDO​​ ​IL​  oad​​  + ​I​ Control​​

where VOUT,LDO and VIN,LDO are the output and input

voltages from the LDO, ILoad is the load c­ urrent drawn
from the LDO, and IControl is the controller current con-
Figure 18. (a) Analog and (b) digital LDO [123]. sumption. It has been shown in [124] that a digital
LDO can offer fast switching at low controller currents,
using the available modes (proximity communi- apart from being synthesizable and process/voltage
cation for mm-scale, IR-HBC for body-connected scalable as compared to its analog counterpart.
devices, low-power wireless/MedRadio for short
distances within 5 m, ANT/BTLE for distances up Dynamically reconfigurable power
to tens of meters, and high-power LoRa WAN for ­conversion LDO
­distances >100 m). To meet the requirements of large DR and high effi-
ciency, which effectively remains constant over the DR
Intelligent energy management of operation, a reconfigurable digital LDO with sam-
By exploiting aggressive technology scaling pling rate adaptation was demonstrated in [124] and
and low-power design techniques, IoT nodes are [125], using an IBM 130-nm CMOS technology. The
continuously trying to reduce the overall energy
­ design comprises of a 128-bit barrel shifter, controlling
consumption. Context-aware and adaptive future
­ 128 identical P-MOSFETs that provide line and load
IoT devices would require different power supplies regulation at the output side (Figure 19a). A clocked,
for different modes of operation, thereby necessitat- sense-amplifier-based comparator compares the reg-
ing an adaptive energy management unit to handle ulated voltage with a reference voltage and, depend-
different scenarios. For example, in-sensor analytics ing on the result of the comparison, increases or
(which are performed in a digital on-chip/on-board decreases the number of P-MOSFETs supplying cur-
processor) can utilize minimal supply voltages for rent to the load. It has been explained in [125] that
low-power, near-threshold (≈0.35 V for 14-nm or lower a linearized control loop model for the LDO has two
nodes) operation, while the high-power RF Tx for long-­ open loop poles—one at DC (from the integrator) and
distance communication may need a higher supply the other at a frequency determined by the ratio of
voltage (≈1 V) for high-power delivery. Short-range the load’s pole frequency to the sampling frequency
communication, on the other hand, may require a dif- of the controller. As the load current dynamically
ferent supply voltage that falls between the minimum changes in different IoT scenarios, the pole from the
and maximum values. Some of these modules may load and, hence, the second pole for the LDO would
be connected to the same power delivery ­network also change in a baseline design without adaptation,
(PDN) (e.g., the entire communication module with leading to overdamped behavior in heavy-load condi-
multiple modalities can be supported by only one tion, and oscillatory behavior in light-load condition,
on-chip voltage regulator), thereby offering a variable which reduces the current efficiency drastically. How-
load to the power management unit (PMU). The PMU ever, if the sampling frequency of the controller is also
may even be supported by a dynamic source if oppor- modified according to the load current (this informa-
tunistic energy harvesting is used to supplement the tion is obtained from how many P-MOSFETs are on in
onboard battery [123]. Hence, the PDN in the con- the LDO), it has been shown that an effectively con-
text-aware, adaptive RC-IoT scenario should be able stant current efficiency can be maintained. This tech-
to support 1) a wide DR, 2) high-power conversion nique for adaptation has two significant system-level
efficiency throughout the range, and 3) a platform advantages: 1) the closed-loop system poles are
and interfacing circuitry for optimum power transfer bounded within a certain range, leading to stable
to the load with minimum losses. LDOs have been tra- and consistent system behavior over a large DR, and
ditionally used in CMOS ASICs to provide ripple-free 2) as the sampling frequency is lowered for light-load

22 IEEE Design&Test

Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
condition, the digital controller’s power also scales,
leading to an improved current efficiency (Figure 19c).
The overall design achieves >80% peak-current
efficiency over 0­.45–1.14 V, with ­ 0.1–4.60-mA load
­current range.

Switched mode control (SMC) for LDOs

To achieve 1) fast transient response across a wide
current and voltage range, 2) rapid droop mitigation,
and 3) dynamic switching by exploiting decoupling
between small signal (SS) gain and large-signal (LS)
transient behavior, Nasir et al. [126] demonstrated an
SMC-based hybrid LDO (analog LDO with SS control
and digital LDO with large signal control) as shown
in Figure 20. SMC combines the advantages of both
analog LDO (high gain, low droop, high-power sup-
ply rejection) with digital LDO (fast LS operation and
adaptivity as shown in Figure 19c) and achieves >80%
peak current efficiency over 0.5–1.1 V, with 1–12-mA
load current and only 6-ns droop recovery time.

Energy management for intermittently Figure 19. (a) Digital LDO with autonomous
powered devices adaptation of sampling CLK that offers a wide DR
Since an IoT device can employ intermittent [124]. (b) Chip micrograph. (c) Current efficiencies
sensing, computation, and communication, which with and without adaptation.
is s­upported from small-energy sources (or from
harvested energy in extremely resource-constrained dynamically and intelligently chosen depending
scenarios), energy management considerations for on information content, accuracy targets, and wire-
intermittent operation become extremely important, less channel conditions. The computing platform,
and lightweight software procedures for control flow,
optimal checkpointing, concurrence, and data con-
sistency [127], [128] need to be developed. This along
with improved techniques of high-dynamic-range,
adaptive PDNs is believed to be one of the major
research directions for context-aware RC-IoT devices.

Intelligent cross-layer
adaptive systems
System-level IoT designs can incorporate more
than one approach discussed previously to optimally
enhance machine intelligence and achieve perfor-
mance improvement/energy reduction, as shown in
[129] and [130].
Cao et al. [129] proposed a camera-based ­wireless
sensor node with a self-optimizing end-to-end com-
putation and communication design, targeted for Figure 20. (a) Hybrid LDO with SMC for wide DR
surveillance applications. The demonstrated system [126]. (b) Chip micrograph. (c) Current efficiencies
supports multiple feature-extraction and classifi- with and without adaptation of sampling frequency
cation algorithms, tunable processing depth (PD), (Fs). SMC(R) is an SMC mode with reset for faster
and PA gain. Minimum-energy ­operating point is droop recovery.

March/April 2019
23
Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 SI: Survey

the consecutive frame differences are computed

as a preprocessing step. Low-power preprocessing
not only locates and segments the potential human
objects but also enables adaptive in situ processing
depending on the information content defined as a
number of active pixels of frame difference. The plat-
form demonstrates three different algorithms to pro-
vide accuracy/energy scalability and each algorithm
is further divided into three PDs—i.e., compression,
feature extraction, and ­ classification—to support
Figure 21. Algorithm demonstration with a real computation and communication tradeoff. On the
video frame [129]. transmission side, the platform applies adaptive
radio whose PA gain is dynamically adjusted, guar-
sensor module, and transmission blocks are ADI anteeing minimum required BER with respect to
ADSP-BF707 image processor, OV7670 camera sen- time-varying wireless channel condition (pathloss
sor, and USRP B200 SDR ­ software-defined radio, in this case). It is intuitive to understand that as the
respectively, and achieve a 4.3× reduction in energy PD increases, the energy cost to compute increases,
consumption compared to a baseline design. but the data volume required to transmit decreases,
The sensing+computation process is demon- thus reducing the energy cost to communicate. As
strated in Figure 21. After image data acquisition, the channel condition changes (from clean to noisy
channel), the minimum energy point also changes.
For a clean channel, a lower PD is preferred (as the
energy to communicate is low), whereas with increas-
ing pathloss a higher PD is preferred. The end-to-end
self-optimization, which dynamically adapts the PD
depending on the channel condition to always track
the point of minimum total energy, is measured and
demonstrated in Figure 22.
Platform proposed in [130] is a collaborative
intelligent heating, ventilation, and air-­conditioning
system (HVAC) occupancy detection solution via
data fusion between optical (OP) and infrared (IR)
camera-based sensor nodes together in a smart wire-
less sensor network. Figure 23 demonstrates that data
fusion has enabled accurate human detection com-
pared with baseline designs, especially in severe light-
ing/heating conditions, and the consequent low miss
rate (5×) in turn reduces sampling rate, resulting in
expanded sensor lifetime (3×) while maintaining the
required detection latency.
Collaborative intelligence is achieved at the sensor
node as well as among the sensor nodes at the back-
end server, which is located at the HVAC and controls
Figure 22. Measured total energy (computation+ the HVAC. When detection accuracy is fixed after the
communication) per frame for different PD with employment of the sensor network, minimizing the
increasing pathloss [129]. Experimental results are latency of occupancy detection depends on reduc-
demonstrated for the three algorithms described ing the sample interval (i.e., the number of OP and IR
here and two BER targets. When pathloss is high, images captured per second). However, a high sam-
the general trend is that optimal mode moves to pling rate will lead to severe sensor energy expend-
more FE-embedded processing. iture and limited sensor lifetime. It is also noted that

24 IEEE Design&Test

Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
the occupancy of a particular region in a building
is dependent on its neighboring regions. For exam-
ple, consider a typical floor-plan of a building with
three rooms, A, B, and C. The occupancy of room A is
dependent on room B if a door between A and B is
available and people can walk from B to A, as shown
in Figure 24a and vice versa. This motivates the pro-
posed dynamic HVAC control strategy, targeting mini-
mized latency of occupancy detection based on a col-
laborative scheme among neighboring HVAC sections.
Consider a network of sensors deployed as shown
in Figure 24a. The sensor node at B estimates the Figure 23. Demonstration of the algorithm
presence of an occupant. If an occupant is detected, presented in [130].
then it further tracks the occupant via difference of
in between the network layer and application layer,
frames and estimation of the d ­ irection of motion.
which is dedicated toward security features and
The direction of motion is sent to the backend,
performs authentication using preshared secrets,
which resolves the potential adjoining HVAC areas
keys, and passwords. However, this layer can also suf-
that can be subsequently occupied. In this example,
fer from DoS attacks and malicious insider attack, as
an occupant moving from B toward A will allow the
illustrated in [137], [155], and [156]. Moreover, the big-
backend to send an alert to the sensor node at A.
data problem in IoT (network exhaustion due to inun-
Now, this sensor node increases its sampling rate to
dation of data) has resulted in modem IoT architects
reduce the latency of detection. The effective sam- to move to a five-layer architecture with added secu-
pling interval, Teff, is reduced as shown in Figure 24b. rity and data-processing capabilities [3], [157]–[163].
References [129] and [130] aim at adaptively CISCO currently defines a seven-layer IoT structure
minimizing energy expenditure of IoT devices in as shown in [135]. In this discussion, we shall limit
a time-varying environment (wireless condition, ourselves to only the perception layer, and hence, a
object moving direction, etc.) while maintaining detailed description of the advanced layer models
decent performance (accuracy, BER, detection (4, 5, and 7 layers) is out of the scope of this article.
latency, etc.) through distributed control on the fly
or centralized control at the backend.
Traditional techniques against
perception layer attacks
Security considerations for A large number of security breaches in RC-IoT
RC-IoT devices occur in the perception layer which is most vul-
From an implementation point of view, the IoT nerable to privacy attacks due to its resource
architecture is usually divided into 3, 4, 5 or 7 ­layers constraints. The most common security measures
as shown in [131]–[135]. References [132]–[134] against attacks on the perception layer are listed in
demonstrated the three-layer architecture as shown
in Figure 25. The details of the three layers and
their security concerns are presented in Table 4.
These security concerns involve data confidential-
ity, integrity, and availability (commonly known as
the CIA triad [136]), which are related to privacy,
correctness, and authentication, respectively. Con-
strained IoT devices (most notably, CO devices such
as small biosensors) have limited resources and can,
hence, support only a subset of the intended secu-
rity features. This makes these devices extremely
prone to privacy attacks [136]–[138]. Figure 24. Illustrative representation of (a) simple
In addition to these three layers, today’s IoT sensor network with interdependence and
devices employ a fourth layer called support layer, (b) demonstration of event-driven sampling [130].

March/April 2019
25
Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 SI: Survey

for practical applications with 14 rounds is still signif-

icantly high with traditional mathematical attacks.
However, nontraditional techniques such as electro-
magnetic side channel attacks (EM-SCA) or power
SCA have proven to reduce the recovery complexity
to a mere 213, breakable within 50 s as shown in [166].
These emerging SCA techniques exploit information
leaks from the physical cryptohardware in the form of
Figure 25. Basic three-layer architecture of IoT EM emanation (for EM-SCA) and supply-line fluctua-
ecosystem. The perception layer devices are tions (for power SCA), and pose a much bigger threat
usually resource constrained and is more prone to to the current standards than computational attacks,
attacks due to elementary security features [131]. motivating the need for hardware techniques to sup-
press the side-channel leaks in the physical system.
1) Power SCA and its countermeasures: Corre-
Table 5. A complete description of each of these
lation power analysis (CPA) [167] has shown to be
techniques is out of the scope of this article. Further
an efficient technique for power SCA as it reduces
details on each can be found in [137] and the cor-
the search space of AES-128/192/256 to just 28 = 256
responding references.
for each key byte (hence, the overall complexity
­becomes 213 for 256-bit, i.e., 25 byte keys). Traditional
Hardware security
power SCA countermeasures try to reduce the SNR of
Encryption engines are at center stage for
the leaked information through power balancing or
achieving IoT security. The computationally secure
gate-level masking but incur significant area, power,
256-bit Advanced Encryption Standard (AES) is tra-
and performance overhead [168]. Attenuated Signa-
ditionally believed to provide data confidentiality
ture Noise Injection (ASNI) was utilized in [168] and
through encryption as the mathematical complex-
[169] (Figure 26), which obfuscates the AES power
ity of the recovery algorithm becomes 2256. AES-256
traces through parallel noise injection and performs
with 14 rounds is used today in banking, military, and
signature attenuation through a signature-attenuating
government applications, and hence, there has been
hardware implemented using an on-chip shunt LDO.
significant research efforts to break (as well as to
This method attenuates the critical AES signature
enhance) the standard. Related subkey-based recov-
in the supply current by >200× with 60% ­additional
ery attacks have shown to significantly decrease the
­overhead in area and 68% overhead in power. More
complexity of key recovery for reduced-round AES
recently, Kar et al. [170] and Singh et al. [171] demon-
(222 for five rounds [164], 239 for nine rounds, and 245
strated the signature attenuation technique using
for ten rounds [165]). However, recovery complexity
an integrated voltage regulator (IVR) and loop ran-
domizing/random fast voltage dithering techniques
with only ≈5% overhead in power and area.
2) EM SCA and its countermeasures: Except for
hardware masking [172], the amount of protective
approaches against correlation EM analysis (CEMA
[173])-based EM-SCA has remained relatively scarce.
In [174], a ground-up approach was presented to find
the specific source of EM emanation within the met-
al stack of an ASIC built using Intel’s 32-nm technolo-
gy, and was generalized using other popular technol-
ogies such as Taiwan Semiconductor Manufacturing
Company (TSMC) 65-nm. It was shown that EM ema-
nations from metals lower than layer eight are barely
distinguishable using a commercially available EM
Figure 26. Noise injection and ASNI as probe, and hence, it was proposed (STELLAR) that
countermeasures for power SCA. a cryptographic core, with power supply routed

26 IEEE Design&Test

Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 Table 4. Details and security considerations of the three-layer IoT architecture [137].

Layer Purpose Security threats Remarks

Perception/sensor Collecting information from sensors/ Eavesdropping [137], Node Most attacks are on
devices Capture [139], Add Malicious Node [137], data confidentiality and
Replay Attack [140], Timing Attack [141] integrity [137], [138]

Network/transmission Connects devices to each Denial of Service (DoS) [137], Most attacks are on data
other and to higher layer through Man-in-the-Middle Attack [142], Storage integrity [137]
wired/wireless media Attack [137], Exploit Attack [137]

Application Has the responsibility to Cross-site Scripting [137], Malicious Code Most attacks are on data
extend sensor-specific services to Attack [137] availability [137]
applications/clients

Table 5. Security measures against perception layer attacks.

Security measure Details Advantages Drawbacks

HMAC [143], [144] Hash Functions along with Encryp- Employed to maintain data Key-hacking is possible through invasive/
tion Algorithms (SHA, MD5, CBC etc) integrity semi-invasive/software/side-channel
are used attacks
Public Key Base station communicates with the More secure than passwords —Ma- 1) Key hacking: The private key needs to
Infrastructure (PKI) devices to get the public key while licious user needs both the secret be protected and 2) not very scalable
protocols [145], [146] the private keys are stored separately private key and a passphrase to
pose any threat
Open Authentication Client-server-based system where Access is granted in a secure way 1) Vulnerable to cross-cite-recovery-
(OAuth/OAuth 2.0) server has the list of authorized forgery (CSRF) and
[147]–[149] clients. Everyone can request for ac- 2) implementation becomes cumbersome
cess, but server grants access tokens as the network grows since the user needs
only to authorized clients to authenticate each device
Mutual authentication Client-server-based system where Both client and server certificates Requires a PKI with high cost of initial
[150], [151] Client creates a request and an are verified deployment
HMAC-SHA signature, and sends
both the request and signature
to server. The server retrieves the
HMAC-SHA signature using a secret
access key and verifies the signature
with client’s signature
Lightweight cryptogra- Cryptographic Keys are used to Plain text to cipher text by using Hard to implement for Class-0 devices with
phy [131], [152] convert messages symmetric, asymmetric keys and stringent resource constraints
hash functions
Embedded security Provides secure secondary storage, Provides a complete security Extremely resource-intensive
framework [153], [154] runtime environment and secure package
memory management

­ ntirely in lower level metals locally, and equipped

e using training data that contains augmented power
with signature suppression techniques like ASNI, be- traces from multiple devices with AES-128. With
fore reaching higher-level metal routing, would be ≈200k traces and proper choice of hyperparam-
resistant against both EM and Power SCA. eters, it was shown that X-DeepSCA attacks can
3) Machine learning SCA—X-DeepSCA and pos- recover keys with 99.9% accuracy from different
sible countermeasures: Recently, ML SCA attacks target devices with ≈10× lower minimum number
have been shown as a big threat as it can uncover of traces as compared to traditional CPA. This in-
the secret key within a few traces using previously creases the threat surface of SCAs significantly and
learned models. Das et al. [175] demonstrate cross- puts further emphasis on SCA countermeasures
device deep-learning-based SCA (X-DeepSCA) such as IVR, ASNI, and STELLAR.

March/April 2019
27
Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 SI: Survey

Arbiters [180], SRAMs [181]–[183], and dynamic

RAMs [184] for PUF implementation, which consume
much less power and area than key-based cryptograph-
ic ­implementations wherein the secret key is stored
in a battery-backed SRAM or in a nonvolatile mem-
ory/electrically erasable programmable read-only
memories (EEPROMs), which are all expensive
­resources for a constrained devices such as C0 or C1.
Figure 27. 128-bit Arbiter PUF [176]. Assuming Moreover, any invasive tampering mechanism
­
same layout length, the circuit creates two delay ­usually changes the PUF’s output, thereby letting the
paths for each input X, and produces the 1-bit out- user know about the attack. These advantages, cou-
put Y based on which path is faster. pled with low resource requirements, makes PUFs a
suitable choice for IoT environments. As an exam-
PUF-based techniques ple, an Arbiter PUF is shown in Figure 27, wherein a
Physical unclonable functions (PUFs) have 128-bit challenge (input) produces a 1-bit response
emerged as a promising augmentation (sometimes (output) according to the respective path delays in
even as an alternative to key/token-based cryp- the data and clock paths due to the random man-
tography), which leverage manufacturing process ufacturing variations [176], [180], and hence, can
variations to generate a unique and device-specific be utilized for device authentication. Even though it
identity for a physical system [176][178]. PUF imple- was shown later in [185] that the randomness of the
mentations are simpler in terms of hardware and output can be modeled with reasonably low com-
do not need to store the secret key that is used to plexity, an improved design with xored outputs from
employ complex cryptographic algorithms. multiple Arbiter PUFs demonstrated high tolerance
1) Digital PUFs: Traditional digital PUFs em- against modeling attacks [185], [186].
ploy simple circuitry such as ring oscillators [179], 2) RF-PUF: Traditional PUF designs discussed
above still require a minimal amount of additional
hardware at the transmitter side of the RC-IoT de-
vice. Chatterjee et al. [187] proposed a new kind of
PUF for RC-IoT scenario, which exploits the effects
of inherent analog and RF process variations at the
transmitter (Tx) side by detecting them with an in-situ
ML hardware at the resource-rich receiver (Rx). This
method embraces the already existing nonidealities
at the Tx, which are usually discarded in a traditional
communication scenario and, hence, do not require
any additional hardware for PUF generation. The
method is inspired by the inherent authentication in
human voice communication as shown in ­Figure 28,
with unique human voice being replaced by unique
Tx signatures, and human brain replaced by a neu-
Figure 28. Principle of RF-PUF [187]. ral network at the Rx. The holistic system-level view
(a) Authentication in human voice communication: for RF-PUF i­mplementation is shown in Figure 29,
Bob (the receiver) can identify Alice (the transmit- while the number of unique transmitters that can be
ter) based on the unique voice signatures, and not identified with varying channel conditions and Rx
based on the contents of what Alice speaks. signatures is shown in Figure 30. It has been shown
Mallory (the impersonator) can also be identified (as with the simulation results that up to 8000 RC-IoT de-
not Alice), since his unique voice signatures would vices can be uniquely identified with 99% accuracy.
be different from Alice. (b) Analogous system that Proof-of-­
concept hardware evaluations were also
utilizes an RF-PUF framework for secure radio ­demonstrated. Since this method does not require any
communication. additional hardware at the Tx, the framework can be

28 IEEE Design&Test

Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
utilized as an extremely useful security feature for RC-
IoT devices for a small-to-medium-scale smart ­system.

Learning frameworks for

RC-IoT ­devices
With the above background in energy-­constrained
sensing, computing, communication, energy manage-
ment, and security, let us look into the ML models that
have been developed and can effectively combine
different modalities based on context. Since this is a
developing field, there is no single learning framework
that overcomes all the challenges in IoT, and hence, it Figure 29. Visualization of RF-PUF in an asymmet-
is important to characterize the most promising learn- ric IoT network with multiple RC-IoT devices as
ing frameworks based on the applications [5]. The transmitters and one resource-rich receiver [187].
current literature on learning in an IoT framework can
broadly be classified into three categories: ML, sequen- and is more suited to RC-IoT devices. It was shown
tial learning (SL), and reinforcement learning (RL). in [190] and [192] that it is possible to converge to
an accurate underlying state using only two previ-
Machine learning (ML) ous states (the tradeoff being higher convergence
ML techniques usually build regression-based time and, hence, more latency, the allowable limit
models on labeled or unlabeled data (for supervised for which depends on the application). Also, unlike
and unsupervised learning, respectively). ML tech- the traditional centralized ML architecture, SL can
niques are computationally complex and require have a distributed implementation and does not
an extensive training data set for acceptable perfor- require an extensive data set for learning. However,
mance, both of which require expensive resources SL requires machine-to-machine (M2M) communi-
[188]. Hence, instead of executing the ML algorithms cation, which may increase energy consumption if
in the RC-IoT device, many of the implementations it is not taken care of at the network implementa-
resort to a centralized cloud-based processing unit tion level. SL is particularly useful for event/anomaly
for ML [3], [188], [189]. However, this means that the detection applications, whereas ML is more suitable
sensor data have to be communicated to the cloud for data analytics applications with higher complex-
for further processing and, hence, pose a burden of ity and higher resource requirement.
communication payload on the RC-IoT device. Com-
pressive sensing and PCA has been shown to be use-
ful [189] in reducing the payload.

Sequential learning (SL)

SL [190]–[192] uses intelligent distributed agents
(RC-IoT devices) that sequentially learn about
an underlying binary state of the system (such as
a medical status, fire alarm, triggering event, or
anomaly- and event-based transmission), and subse-
quently propagate it through the network, as shown
in Figure 31. Depending on the number of previous
agents from which information is gathered, SL is cat-
egorized into finite memory and infinite memory [5].
In infinite memory SL, agents collect information on
the estimate from all other agents in the sequence Figure 30. Probability of false detection
and, hence, require more memory resources. Finite as a function of the total number of trans-
memory SL, on the other hand, collects information mitters in the system, with and without
from a user-­defined fixed number of previous agents receiver signature compensation [187].

March/April 2019
29
Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 SI: Survey

based on the state transition due to the action taken

at a given state. The total cumulative reward (called
Q value) of performing an action at at a given state
st is given by the linear combination of the old Q
value, the immediate reward and the total estimate
of the future rewards as indicated by the Q-learning-­
equation in Figure 32a.
Amravati et al. [197] demonstrated a time-
domain mixed-signal neuromorphic accelerator with
Figure 31. Sequential learning in IoT [5]. embedded RL implemented using a three-layer neu-
ral network with 84 neurons. The test chip was built
in 55-nm CMOS technology and was mounted on
Reinforcement learning (RL)
a mobile microrobot for autonomous exploration
RL implementations [188], [193]–[197] utilize
of the environment (Figure 32b). The peak power
the interaction between the agent and the environ-
was only 690 µW at 1.2-V supply while operating at
ment in a method based on rewards and penalties.
3.12 TOPS/W. The peak energy efficiency was 690 pJ/
In RL, the agents can perform a predefined set of
Inference and 1.5 nJ/training (1.25 pJ/MAC), making
actions in an environment with a set of states and
it one of the highest performance and lowest power
a state-transition function. The action of the agents
implementation till date. The low energy is attributed
changes the states, and based on the state transi-
to: 1) time-domain mixed-signal MAC operations with
tions and the final goal, the environment rewards
time-domain inputs which do not need voltage to time
(or penalizes) the agent. The agent tries to maximize
or time to voltage conversions and consumes scaled
its immediate as well as future rewards and learns
energies based on the importance of the computa-
to converge to a steady state as explained in [193].
tion and 2) a relatively low 6-bit precision, which was
The action-reward combination works as a closed
shown to be enough for l­ow-to-medium complexity
loop feedback system with a high chance of conver-
applications [45] involving pattern/object recognition.
gence, and can be implemented using computation-
The convergence process for RL is slower than
ally simple algebraic Q-learning algorithm [188]. As
SL and the requirement to preemptively know the
shown in Figure 32a, the agent receives the reward
states and state-transition-matrix makes RL chal-
lenging for medium- and high-complexity applica-
tions. However, for low-complexity, high-latency-­
tolerant tasks such as resource management or
power ­ management [195], [196], RL can be an
extremely relevant choice. Parallel Q-learning (PQL)
algorithms (PCSP-8 [198], PQL-C [199], CS-RL, and
­CS-RL-EXT [200]) have also been developed for dis-
tributed, resource-constrained applications and for
speeding up the RL convergence. Figure 33 shows a
comparison of speed-up using these techniques.
In essence, SL- and RL-based learning techniques
have shown enough promise through lightweight
algorithms that can be implemented on the small
IoT nodes. However, a network-wide full realization
of these techniques for context discovery and assess-
ment is still a wide open area of research. New devices
and technologies such as mixed-signal neurons,
Figure 32. (a) Q-learning (RL framework) for RC-IoT memristors, spin-transfer-torque-based devices, opto-
devices. (b) Implementation showing an RL-ena- electronic and ferroelectric devices with in-memory
bled microrobot [197] with time-domain inputs and and near-memory computation to reduce memory
processing through a three-layer neural network fetch, computation, and communication power in a
and SVM loss function. neural network are areas of active research and hold

30 IEEE Design&Test

Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
tremendous potential for future. Online/incremental
learning is of paramount importance because of the
variations in the manufacturing process and oper-
ating conditions. By fully utilizing the capabilities of
devices, hardware, and algorithms together, the path
toward more efficient context-aware systems needs to
be paved.

The way ahead: Future of IoT systems

Where we stand: Systems with efficient
components and promising applications
The preceding analysis puts us in perspective of
the current state-of-the-art in adaptive, context-aware Figure 33. Convergence speed-up using parallel
IoT hardware and machine intelligence. The number Q-learning algorithms (PQL with coallocation of
of applications is numerous, ranging from small-scale storage and processing: PCSP-8 [198], PQL with
smart biosensing and smart cars to medium-scale local cache: PQL-C [199], constant share RL: CS-
smart homes and offices, and to large-scale smart RL, and extended CS-RL: CS-RL-EXT [200]).
cities. Significant research efforts have been put into
optimizing the available resources for correspond- reduce subsequent power consumption in com-
ing applications as pointed out in p ­ revious sections. municating otherwise raw data bits to the cloud.
Anomaly/Outlier/Event detection and data
Where the future lies: Secure, context-aware, compression are the two most important forms
intelligent, and adaptive devices and systems of in-sensor/edge analytics that are required in
However, bigger IoT networks still remain subop- today’s systems and is an extremely promising
timal in terms of effective utilization of the distrib- research direction for bringing down the power
uted resources. The vision of secure, context-aware, consumption due to nonoptimal data handling
intelligent, and adaptive devices and systems, as pre- and communication.
sented throughout this article, involves a holistic opti- • Communication: As shown in the analysis pre-
mization of all the resource-constrained leaf devices sented in the “Intelligent Computing Platforms”
within a network, in each of the following subareas. and the Intelligent Communication” sections, this
is the subsystem toward which a lot of research
• Sensing: As the sensing leaf nodes in an IoT focus should be directed for practical feasibil-
network have the most stringent resource con- ity of the context-aware vision. The numerous
straints, the sensing process itself should be made modalities available (proximity communication
extremely low power through sub-Nyquist-rate CS [83], HBC [84], NFC, ZigBee, ANT, BTLE, Wi-Fi
[22], [30], [33] for sparse signals (such as audio and LoRA, among others) makes this a multidi-
and image), or made adaptive/energy-­resolution mensional and multilevel optimization problem
scalable through time-/frequency-based sens- with possibilities of intra-PHY and inter-PHY
ing [40] for slowly varying signals with high DR adaptability and tradeoffs. Techniques such as
(such as radiation and vibration). The adaptivity anomaly detection and channel quality estima-
information/resolution requirement (context) tion would determine when to communicate,
should come from the cloud for latency-relaxed and how much data are to be sent (e.g., burst-
applications, and from in-senor/on-gateway mode communication will bring in further energy
learning hardware for latency-limited scenarios. efficiency and context awareness on top of data
Reconfigurability among Nyquist-rate sensing, compression, through duty cycled intermittent
CS and time-/frequency-based sensing can be an communication, even with good channel qual-
optional feature, depending on the applications ity [129]). Furthermore, short-range low-energy
and amount of resources available. communication using HBC/ANT/BTLE should be
• Computation: The intelligent RC-IoT nodes explored to assess the possibility of spatial data
should have the capability of locally extracting compression for sensors within close proximity
important information from the sensed data to of each other. If the spatial data compression

March/April 2019
31
Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 SI: Survey

is possible, only one node in an RC-IoT cluster in the IoT environment, it is necessary to model the
would take the responsibility to communicate heterogeneity, resource ­constraints, and distribution
the compressed data to the upper level gateway/ of the IoT devices within the architecture in a struc-
cloud (possibly using a higher power modal- tured manner. CHT is an emerging tool to provide
ity like LoRa for long-range communication). such a modeling framework using behavioral game
Again, processing all the above information theory, and is based on bounded rationalities [5],
would require sophisticated learning algorithms [201], [202]. The theory of bounded rationalities
to be implemented in different hierarchical lev- ensures that each node in the network tries to find its
els of the IoT architecture which, by itself, is an best strategy, bounded by information from the lower
involved optimization problem. level nodes in the hierarchy, its own computational
• Energy management: As shown in the “Intelligent capacity, and time/resource available. CHT model
Energy Management” section, high-dynamic-­ (Figure 34) inherently takes care of the device het-
range and high-power conversion LDOs with low erogeneity in IoT as it considers the resources availa-
voltage droop/droop recovery time are one of the ble for each device separately. References [201] and
major requirements in a dynamic IoT scenario. [202] present further details of the CHT techniques,
High DR adaptation techniques such as sampling while [5] demonstrates an example of the CHT theory
frequency based reconfigurable LDOs [124] and in determining the type of learning algorithm (ML, SL,
SMC LDOs [126] have been explored. However, and RL) to be implemented on a particular IoT device
challenges due to checkpointing and data con- based on its resource constraints. It must be noted
sistency need to be looked into. Recent check- that though CHT would define a structure in the het-
pointing schemes such as the one shown in [128] erogeneous IoT hierarchy, such an algorithm cannot
have demonstrated improved latencies in a medi-
be implemented in C0 and possibly C1 devices. How-
um-to-high-resource device—though similar and
ever, the output of the algorithm can be passed on
more lightweight techniques need to be devel-
to the RC-IoT devices from higher level nodes which
oped for highly resource-constrained devices.
have higher computational power.
• Adaptive security: The RF-PUF [187] framework
shown in the “Learning Frameworks for RC-IoT IoT networks are different from traditional net-
Devices” section, along with low-level metal rout- works in view of their specific challenges in device
ing for the encryption core [174] for EM-SCA heterogeneity, resource constraints, context-variability,
resistance can be utilized as a baseline security and security, thereby necessitating adaptive solutions
feature at no additional power/area overhead in for resource-aware operation. In this article, we have
extremely resource-constrained C0 devices. Light- presented a broad review of the different areas that
weight implementation of ASNI/IVR [168], [170] need to be looked into for holistic, system-level
with minimal overhead should be placed as well resource optimization for RC-IoT devices in a network.
in C0 devices, while nodes with more relaxed con- Various techniques in sensing (compressed-domain
straints can benefit from implementations with sensing/energy-resolution scalable frequency-domain
better signature attenuation (consum-
ing higher power). These techniques
should also be augmented with one or
multiple traditional security features
such as hash-based message authen-
tication (HMAC) and mutual authen-
tication/OAuth based on the context
(application, importance of collected
data) and resources available, and can
be adaptive in nature.

Cognitive hierarchy theory (CHT)

To capture and exploit the multitude Figure 34. Distribution of IoT devices according to
of reconfigurable modalities effectively the levels of cognitive hierarchy theory [5].

32 IEEE Design&Test

Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
sensing), computation (in-sensor/edge analytics in the [7] C. Perera et al., “A survey on Internet of Things from
form of outlier detection and data compression), com- industrial market perspective,” IEEE Access, vol. 2,
munication (intra-PHY and inter-PHY adaptation with pp. 1660–1679, 2014.
low and high-power modalities), power management [8] The ‘Only’ Coke Machine on the Internet.
and security were analyzed, and the vision for a secure, Accessed: December 9, 2018 [Online]. Available:
context-aware, adaptive, resource-constrained but https://www.cs.cmu.edu/~coke/history_long.txt
intelligent IoT device was presented. However, numer- [9] M. Weiser, “The computer for the 21st century,”
ous challenges (in the form of system-level controller Scientific American, Sep. 1991.
design for adaptive architectures, reliability, security, [10] K. Ashton, That ‘Internet of Things’ Thing. Accessed:
latency limitations, intermittent ­powering/checkpoint- December 9, 2018. [Online]. Available: https://www.
ing and real-time/online learning) still exist in realizing rfidjournal.com/articles/view?4986
a full implementation of the concepts demonstrated, [11] P. Langley and H. A. Simon, “Applications of machine
indicating future research directions toward building learning and rule induction,” Commun. ACM, vol. 38,
smarter and more adaptive systems. In that context, pp. 54–64, Nov. 1995.
the goal in this article has been to identify the current [12] A. S. Tanenbaum and A. S. Woodhull, Operating Syst.
trends, foundations and components of the envisioned Design and Implementation, Upper Saddle River,
RC-IoT devices to ­enable the design of more efficient NJ: Prentice-Hall, 2005, 3rd ed.
connected intelligent systems in the future.  [13] C. Bormann, M. Ersue, and A. Keranen, Terminology
for Constrained-node Netw., RFC 7228. Accessed:
Acknowledgments January 21, 2019. [Online]. Available: https://www.rfc-
This work was supported in part by the Semicon- editor.org/info/rfc7228
ductor Research Corporation (SRC) under Grant [14] B. N. Schilit and M. M. Theimer, “Disseminating active
2720.001 and in part by the National Science Foun- map information to mobile hosts,” IEEE Netw., vol. 8,
dation (NSF) CRII Award under Grant CNS 1657455. pp. 22–32, Sep. 1994.
[15] G. D. Abowd et al., “Toward a better understanding
 References of context and context-awareness,” in Proc. 1st Int.
[1] R. Lloyd, Mobile Traffic From Wearables Explodes Symp. Handheld Ubiquitous Comput.’99, London, UK:
as the Internet of Everything Accelerates. Accessed: Springer-Verlag, 1999, pp. 304–307.
December 5, 2018. [Online]. Available: http://blogs. [16] A. K. Dey, G. D. Abowd, and D. Salber, “A conceptual
cisco.com/news/mobile-traffic-from-wearables- framework and a toolkit for supporting the rapid
explodes-as-the-internet-of-everything-accelerates prototyping of context-aware applications,” Hum.-
[2] S. Sen, “Invited: Context-aware energy-efficient Comput. Interact., vol. 16, pp. 97–166, Dec. 2001.
communication for IoT sensor nodes,” in Proc. 2016 [17] P. J. Brown, “The STICK-e document: A framework for
53nd ACM/EDAC/IEEE Design Autom. Conf., creating context-aware applications,” in Electron. Pub.,
Jun. 2016, pp. 1–6. pp. 259–272, 1996.
[3] J. Gubbi et al., “Internet of Things (IoT): A vision, [18] D. Franklin and J. Flachsbart, All Gadget and No
architectural elements, and future directions,” Future Representation Makes Jack a Dull Environment, 1998.
Gener. Comput. Syst., vol. 29, no. 7, pp. 1645–1660, [19] T. Rodden et al., “Exploiting context in HCI design for
2013. mobile systems,” in Proc. Workshop Hum. Comput.
[4] F. Delicato, P. Pires, and T. Batista, “Resource Interact. Mobile Dev., 1998.
management for Internet of Things,” in The Resource [20] R. Hull, P. Neaves, and J. Bedford-Roberts, “Toward
Management Challenge in IoT, Cham, Switzerland: situated computing,” in Proc. Digest Papers First Int.
Springer, 2017. Symp. Wearable Comp., Oct. 1997, pp. 146–153.
[5] T. Park, N. Abuzainab, and W. Saad, “Learning how [21] K. Henricksen and J. Indulska, “Developing context-
to communicate in the Internet of Things: Finite aware pervasive computing applications: Models and
resources and heterogeneity,” IEEE Access, vol. 4, approach,” Pervasive Mob. Comput., vol. 2, pp. 37–64,
pp. 7063–7073, 2016. Feb. 2006.
[6] C. Perera et al., “Context aware computing for the [22] E. J. Candes and T. Tao, “Decoding by linear
Internet of Things: A survey,” IEEE Commun. Surveys programming,” IEEE Trans. Inf. Theory, vol. 51,
Tuts., vol. 16, pp. 414–454, First 2014. pp. 4203–4215, Dec. 2005.

March/April 2019
33
Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 SI: Survey

[23] D. L. Donoho, “Compressed sensing,” IEEE Trans. Inf. [37] J. D. Roose et al., “Flexible and self-adaptive sense-
Theory, vol. 52, pp. 1289–1306, Apr. 2006. and-compress for submicrowatt always-on sensory
[24] M. F. Duarte et al., “Single-pixel imaging via recording,” in ESSCIRC 2018—IEEE 44th Euro. Solid
compressive sampling,” IEEE Signal Process. Mag., State Circ. Conf., pp. 282–285, Sep. 2018.
vol. 25, pp. 83–91, Mar. 2008. [38] M. Trakimas and S. R. Sonkusale, “An adaptive
[25] M. Lustig, D. Donoho, and J. M. Pauly, “Sparse MRI: resolution asynchronous ADC architecture for
The application of compressed sensing for rapid MR data compression in energy constrained sensing
imaging,” Magn. Reson. Med., vol. 58, no. 6, applications,” IEEE Trans. Circ. Syst. I, vol. 58,
pp. 1182–1195, 2007. pp. 921–934, May 2011.
[26] R. Baraniuk and P. Steeghs, “Compressive radar [39] C. Ieong et al., “A 0.45 V 147-375 nW ECG
imaging,” in Proc. 2007 IEEE Radar Conf., Apr. 2007, compression processor with wavelet shrinkage and
pp. 128–133. adaptive temporal decimation architectures,” IEEE

[27] A. Amaravati et al., “A 130nm 165nJ/frame Trans. Very Large Scale Integr. Syst., vol. 25,

compressed-domain smashed-filter based mixed- pp. 1307–1319, Apr. 2017.

signal classifier for “In-sensor” analytics in smart [40] B. Chatterjee et al., “A wearable real-time CMOS
dosimeter with integrated zero-bias floating-gate
cameras,” Trans. Circ. Syst. II, no. 9, 2017.
sensor and an 861nW 18–bit energy-resolution
[28] A. Anvesha et al., “A light-powered, always-on, smart
scalable time-based radiation to digital converter,” in
camera with compressed domain gesture detection,”
2019 IEEE Custom Integr. Circ. Conf., Apr. 2019.
in Proc. 2016 Int. Symp. Low Power Electron. Design,
[41] W. Li, J. Bao, and W. Shen, “Collaborative wireless
2016, pp. 118–123.
sensor networks: A survey,” in 2011 IEEE Int. Conf.
[29] S. Xu et al., “Appearance-based gesture recognition
Syst. Man Cybern., pp. 2614–2619, Oct. 2011.
in the compressed domain,” in 2017 IEEE Int. Conf.
[42] M. Singh and V. K. Prasanna, “A hierarchical model
Acoust., Speech Signal Process., pp. 1722–1726,
for distributed collaborative computation in wireless
Mar. 2017.
sensor networks,” in Proc. Int. Parallel Distrib.
[30] A. Anvesha et al., “A 65nm compressive-sensing
Process. Symp., Apr. 2003, p. 166.2.
time-based ADC with embedded classification and
[43] R. M. Buehrer, H. Wymeersch, and R. M. Vaghefi,
INL-aware training for arrhythmia detection,” in 2017
“Collaborative sensor network localization: Algorithms
IEEE Biomedical Circ. and Syst. Conf., pp. 1–4,
and practical issues,” in Proc. IEEE, vol. 106,
Oct. 2017.
Jun. 2018, pp. 1089–1114.
[31] S. Qaisar et al., “Compressive sensing: From theory to
[44] N. Cao et al., “In-sensor analytics and energy-aware
applications, a survey,” J. Commun. Netw., vol. 15,
self-optimization in a wireless sensor node,” in Proc.
pp. 443–456, Oct, 2013.
2017 IEEE MTT-S Int. Microw. Symp., pp. 200–203,
[32] H. Djelouat, A. Amira, and F. Bensaali, “Compressive
Jun. 2017.
sensing-based IoT applications: A review,” J. Sensor
[45] B. Chatterjee et al., “Exploiting inherent error-
Actuator Netw., vol. 7, no. 4, 2018. resiliency of deep neural networks to achieve extreme
[33] A. Amaravati et al., “A light-powered smart camera energy-efficiency through mixed-signal neurons,”
with compressed domain gesture detection,” IEEE ArXiv:1806.05141 (CS), vol. abs/1806.05141, 2018.
Trans. Circ. and Syst. for Video Technology, vol. 28, [46] B. Chatterjee et al., “An energy-efficient mixed-signal
pp. 3077–3085, Oct. 2018. neuron for inherently error-resilient neuromorphic
[34] J. Zhang, Z. Wang, and N. Verma, “18.4 A matrix- systems,” in Proc. 2017 IEEE Int. Conf. Reboot.
multiplying ADC implementing a machine-learning Comput., 2017, pp. 1–2.
classifier directly with data conversion,” in Proc. ISSCC [47] C. H. Bennett, “Notes on Landauer’s principle, reversible
2015, Feb. 2015, pp. 1–3. computation, and Maxwell’s Demon,” Studies Hist. Phil.
[35] S. S. Rautaray and A. Agrawal, “Vision based hand Modern Phys., vol. 34, no. 3, pp. 501–510, 2003.
gesture recognition for human computer interaction: A [48] H. T. Friis, “A note on a simple transmission formula,” in
survey,” Artifi. Intell. Rev., vol. 43, no. 1, pp. 1–54, 2015. Proc. IRE, vol. 34, May 1946, pp. 254–256.
[36] V. I. Pavlovic, R. Sharma, and T. S. Huang, “Visual [49] A. J. Johansson, “Performance of a radio link between
interpretation of hand gestures for human-computer a base station and a medical implant utilising the
interaction: A review,” IEEE Trans. Pattern Anal. Mach. MICS standard,” in 26th Ann. Int. Conf. IEEE Eng.
Intell., vol. 19, no. 7, pp. 677–695, 1997. Med. Biol. Soci., vol. 1, pp. 2113–2116, Sep. 2004.

34 IEEE Design&Test

Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
[50] C. Salazar et al., “A 2.4 GHz interferer-resilient [65] P. Malhotra et al., “Long short term memory networks
wake-up receiver using a dual-IF MultiStage N-path for anomaly detection in time series,” in ESANN, 2015.
architecture,” IEEE J. Solid-State Circ., vol. 51, [66] S. Ahmad et al., “Unsupervised realtime anomaly
pp. 2091–2105, Sep. 2016. detection for streaming data,” Neurocomputing,
[51] B. Razavi, RF Microelectronics. Upper Saddle River, vol. 262, pp. 134–147, 2017.
NJ: Prentice Hall, 2011, 2nd ed.. [67] V. Chatzigiannakis et al., “Hierarchical anomaly detection
[52] A. Ebrazeh and P. Mohseni, “30 pJ/b, 67 Mbps, in distributed large-scale sensor networks,” in 11th IEEE
centimeter-to-meter range data telemetry with an Symp. Comput. Commun.’06, pp. 761–767, Jun. 2006.
IR-UWB wireless link,” IEEE Trans. Biomed. Circ. [68] Y. Lee, Y. Yeh, and Y. F. Wang, “Anomaly detection via
Syst., vol. 9, pp. 362–369, Jun. 2015. online oversampling principal component analysis,” IEEE
[53] W. Wu et al., “Localized outlying and boundary data Trans. Knowl. Data Eng., vol. 25, pp. 14601470, Jul. 2013.
detection in sensor networks,” IEEE Trans. Knowl. [69] V. Barnett and T. Lewis, Outliers in Statistical Data.
Data Eng., vol. 19, pp. 1145–1157, Aug. 2007. John Wiley & Sons Ltd., 1994, 2nd ed.
[54] C.-C. J. K. Minwook et al., “Distributed spatiotemporal [70] Y. Zhang, N. Meratnia, and P. Havinga, “Outlier
outlier detection in sensor networks,” in Proc. SPIE, detection techniques for wireless sensor networks:
vol. 5819, 2005, pp. 5819–5819–12. A survey,” IEEE Commun. Surveys Tuts., vol. 12,
[55] T. Palpanas et al., “Distributed deviation detection in pp. 159–170, Second 2010.
sensor networks,” SIGMOD Rec., vol. 32, pp. 77–82, [71] J. Hochenbaum, O. S. Vallis, and A. Kejariwal,
Dec. 2003. “Automatic anomaly detection in the cloud via
[56] S. Papadimitriou et al., “LOCI: Fast outlier detection statistical learning,” ArXiv. 1704.07706, 2017.
using the local correlation integral,” in Proc. 19th Int. [72] K. C. Barr and K. Asanovic, “Energy-aware lossless
Conf. Data Eng., Mar. 2003, pp. 315–326. data compression,” ACM Trans. Comput. Syst.,
[57] B. Sheng, Q. Li, W. Mao, and W. Jin, “Outlier detection vol. 24, pp. 250–291, Aug. 2006.
in sensor networks,” in Proc. 8th ACM Int. Symp. [73] D. Garlan et al., “Project Aura: Toward distraction-free
Mobile Ad Hoc Netw. Comput. MobiHoc ’07, New York, pervasive computing,” IEEE Pervasive Comput., vol. 1,
NY, 2007, pp. 219–228. pp. 22–31, Apr. 2002.
[58] J. Branch et al., “In-network outlier detection in [74] C. Sadler and M. Martonosi, “Data compression
wireless sensor networks,” in 26th IEEE Int. Conf. algorithms for energy-constrained devices in delay
Distrib. Comput. Syst. ’06, pp. 51–51, Jul. 2006. tolerant networks,” in Proc. ACM Conf. Embedded
[59] Y. Zhuang and L. Chen, “In-network outlier cleaning Netw. Sensor Syst., 2006, pp. 265–278.
for data collection in sensor networks,” in In CleanDB [75] A. Rooshenas et al., “Reducing the data transmission in
Workshop VLDB 2006, pp. 41–48, 2006. wireless sensor networks using the principal component
[60] S. Rajasegarar et al., “Distributed anomaly detection in analysis,” in Proc. 2010 Sixth Int. Conf. Intell. Sensors
wireless sensor networks,” in 2006 10th IEEE Singapore Sensor Netw. Inf. Process., Dec. 2010, pp. 133–138.
Int. Conf. Commun. Syst., pp. 1–5, Oct. 2006. [76] A. Morell et al., “Data aggregation and principal
[61] S. Rajasegarar et al., “Quarter sphere based distributed component analysis in WSNs,” IEEE Trans. Wireless
anomaly detection in wireless sensor networks,” in 2007 Commun., vol. 15, pp. 39083919, Jun. 2016.
IEEE Int. Conf. Commun., pp. 3864–3869, Jun. 2007. [77] S. V. Macua, P. Belanovic, and S. Zazo, “Consensus-
[62] N. Shahid, I. H. Naqvi, and S. B. Qaisar, “One-class based distributed principal component analysis
support vector machines: Analysis of outlier detection in wireless sensor networks,” in Proc. 2010 IEEE
for wireless sensor networks in harsh environments,” 11th Int. Workshop Signal Process. Adv. Wireless
Artif. Intell. Rev., vol. 43, pp. 515–563, Apr. 2015. Commun., Jun. 2010, pp. 1–5.
[63] D. Janakiram, V. A. Reddy, and A. V. U. P. Kumar, [78] D. Petrovic et al., “Data funneling: Routing with
“Outlier detection in wireless sensor networks using aggregation and compression for wireless sensor
bayesian belief networks,” in 2006 1st Int. Conf. networks,” in Proc. First IEEE Int. Workshop Sensor
Commun. Syst. Softw. Middlew., pp. 1–6, Jan. 2006. Netw. Protocols Appl., May 2003, pp. 156–162.
[64] E. Elnahrawy and B. Nath, “Context-aware sensors,” in [79] T. Arici et al., “PINCO: A pipelined in-network
Wireless Sensor Netw., EWSN 2004. Lecture Notes in compression scheme for data collection in wireless
Computer Science, Berlin, Heidelberg: Springer, Jan. sensor networks,” in Proc. 12th Int. Conf. Comput.
2004, vol. 2920. Commun. Netw., Oct. 2003, pp. 539–544.

March/April 2019
35
Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 SI: Survey

[80] E. Magli, M. Mancin, and L. Merello, “Low-complexity   [94] P. Malla et al., “A 28mW spectrum-sensing
video compression for wireless sensor networks,” in reconfigurable 20MHz 72dB-SNR 70dB-SNDR DT
Proc. 2003 Int. Conf. Multimedia Expo ’03, vol. 3, AY ADC for 802.11n/WiMAX receivers,” in Proc. 2008
Jul. 2003, pp. III-585. IEEE Int. Solid-State Circ. Conf.—Digest Tech. Pap.,
[81] S. S. Pradhan, J. Kusuma, and K. Ramchandran, Feb. 2008, pp. 496–631.
“Distributed compression in a dense microsensor   [95] Aurangozeb et al., “Channel-adaptive ADC and TDC
network,” IEEE Signal Process. Mag., vol. 19, for 28 Gb/s PAM-4 digital receiver,” IEEE J. Solid-
pp. 51–60, Mar. 2002. State Circ., vol. 53, pp. 772788, Mar. 2018.
[82] S. Sen, J. Koo, and S. Bagchi, “TRIFECTA: Security,   [96] R. Senguttuvan, S. Sen, and A. Chatterjee, “VIZOR:
energy efficiency, and communication capacity Virtually zero margin adaptive RF for ultralow power
comparison for wireless IoT devices,” IEEE Internet wireless communication,” in Proc. 2007 25th Int.
Comput., vol. 22, pp. 74–81, Jan. 2018. Conf. Comput. Design, Oct. 2007, pp. 580–586.
[83] C. Thakkar et al., “A 32 Gb/s bidirectional 4-channel   [97] R. Senguttuvan, S. Sen, and A. Chatterjee,
4 pJ/b capacitively coupled link in 14 nm CMOS for “Multidimensional adaptive power management for
proximity communication,” IEEE J. Solid-State Circ., low-power operation of wireless devices,” IEEE Trans.
vol. 51, pp. 3231–3245, Dec. 2016. Circ. and Syst. 11, vol. 55, pp. 867–871, Sep. 2008.
[84] S. Maity et al., “A 6.3pJ/b 30Mbps -30dB SIR-   [98] S. Sen, M. Verhelst, and A. Chatterjee, “Orthogonally
tolerant broadband interference-robust human body tunable inductorless RF LNA for adaptive wireless
communication transceiver using time domain signal- systems,” in 2011 IEEE Int. Symp. Circ. and Syst.,
interference separation,” in Proc. 2018 IEEE Custom pp. 285–288, May 2011.
Integr. Circ. Conf., Apr. 2018, pp. 1–4.   [99] S. Sen et al., “A power-scalable channel-adaptive
[85] D. Das et al., “Enabling covert body area network wireless receiver based on built-in orthogonally
using electro-quasistatic human body communication,” tunable LNA,” IEEE Trans. Circ. and Syst. I, vol. 59,
in Sci. Rep., Jan. 2019. pp. 946–957, May 2012.
[86] I. Akyildiz et al., “Wireless sensor networks: A survey,” [100] D. Banerjee et al., “Realtime use-aware adaptive
Comput. Netw., vol. 38, no. 4, pp. 393–422, 2002. MIMO RF receiver systems for energy efficiency
[87] A. J. Goldsmith and S. Chua, “Adaptive coded under BER constraints,” in Proc. 2013 50th ACM/
modulation for fading channels,” IEEE Trans. EDAC/IEEE Design Autom. Conf., May 2013, pp. 17.
Commun., vol. 46, pp. 595–602, May 1998. [101] D. Banerjee et al., “Real-time use-aware adaptive
[88] A. Srivastava et al., “Bio-WiTel: A low-power integrated RF transceiver systems for energy efficiency under
wireless telemetry system for healthcare applications BER constraints,” IEEE Trans. Comput.-Aided Design
in 401–406 MHz band of MedRadio spectrum,” IEEE J. Integr. Circ. Syst., vol. 34, pp. 1209–1222, Aug. 2015.
Biomed. Health Inf., vol. 22, pp. 483–494, Mar. 2018. [102] S. Sen et al., “ProVIZOR: Process tunable virtually
[89] S. Sen, R. Senguttuvan, and A. Chatterjee, zero margin low power adaptive RF for wireless
“Environment-adaptive concurrent companding and systems,” in Proc. 2008 45th ACM/IEEE Design
bias control for efficient power-amplifier operation,” IEEE Autom. Conf., Jun. 2008, pp. 492–497.
Trans. Circ. Syst. I, vol. 58, pp. 607–618, Mar. 2011. [103] S. Sen, “Channel-adaptive zero-margin and process-
[90] S. Sen et al., “Process-variation tolerant channel-adaptive adaptive self-healing communication circuits/
virtually zero-margin low-power wireless receiver systems: Special session paper,” in Proc. 2014 IEEE/
systems,” IEEE Trans. Comput.-Aided Design Integr. ACM Int. Conf. Comput.-Aided Design, Nov. 2014,
Circ. and Syst., vol. 33, pp. 1764–1777, Dec. 2014. pp. 80–85.
[91] J. Zhu, H. Krishnaswamy, and P. R. Kinget, “Field- [104] D. Banerjee et al., “Self-learning MIMO-RF receiver
programmable LNAs with interferer-reflecting loop for systems: Process resilient real-time adaptation to
input linearity enhancement,” IEEE J. Solid-State Circ., channel conditions for low power operation,” in Proc.
vol. 50, pp. 556–572, Feb. 2015. 2014 IEEE/ACM Int. Conf. Comput.-Aided Design,
[92] F. Behbahani et al., “Adaptive analog IF signal processor Nov. 2014, pp. 710–717.
for a wide-band CMOS wireless receiver,” IEEE J. Solid- [105] D. Banerjee, S. Sen, and A. Chatterjee, “Self learning
State Circ., vol. 36, pp. 1205–1217, Aug. 2001. analog/mixed-signal/RF systems: Dynamic adaptation
[93] D. T. Lin et al., “A low-power adaptive receiver utilizing to workload and environmental uncertainties,” in Proc.
discrete-time spectrum-sensing,” IEEE Trans. Microw. the IEEE/ACM Int. Conf. Comput.-Aided Design ’15,
Theory Techn., vol. 61, pp. 1338–1346, Mar. 2013. Piscataway, NJ, 2015, pp. 59–64.

36 IEEE Design&Test

Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
[106] D. Banerjee et al., “Self-learning RF receiver systems: [119] J. Jaussi et al., “26.2 a 205mw 32gb/s 3-tap ffe/6-tap
Process aware real-time adaptation to channel dfe bidirectional serial link in 22nm CMOS,” in Proc.
conditions for low power operation,” IEEE Trans. Circ. 2014 IEEE Int. Solid-State Circ. Conf. Digest Techn.
and Syst. 1, vol. 64, pp. 195–207, Jan. 2017. Pap., 2014, pp. 440–441.
[107] S. Sen, “Channel-adaptive zero-margin & process- [120] T.-C. Hsueh et al., “26.4 a 25.6 gb/s differential and
adaptive selfhealing communication circuits/systems: ddr4/gddr5 dual-mode transmitter with digital clock
Special session paper,” in Proc. 2014 IEEE/ACM Int. calibration in 22nm CMOS,” in Proc. 2014 IEEE Int.
Conf. Comput.-Aided Design, 2014, pp. 80–85. Solid-State Circ. Conf. Digest Techn. Pap., 2014,
[108] C. Thakkar et al., “23.2 a 32gb/s bidirectional pp. 444–445.
4-channel 4pj/b capacitively coupled link in 14nm [121] S. Maity, D. Das, and S. Sen, “Adaptive interference
cmos for proximity communication,” in Solid-State rejection in human body communication using
Circ. Conference (ISSCC), 2016 IEEE Int., variable duty cycle integrating DDR receiver,” in Proc.
pp. 400–401, 2016. 2017 Design Autom. Test Euro. Conf. Exhibition,
[109] T. Tsukizawa et al., “A fully integrated 60GHz CMOS 2017, pp. 1763–1768.
transceiver chipset based on WiGig/IEEE802.11ad [122] S. Maity, P. Mehrotra, and S. Sen, “An improved
with built-in self calibration for mobile applications,” in update rate baud rate CDR for integrating human
Proc. ISSCC 2013, Feb. 2013, pp. 230–231.
body communication receiver,” in Proc. 2018 IEEE
[110] M. He et al., “A 40nm dual-band 3-stream 802.11a/b/
Biomedical Circ. and Syst. Conf., 2018, pp. 1–4.
g/n/ac mimo WLAN SoC with 1.1gb/s over-the-air
[123] S. Gangopadhyay, S. B. Nasir, and A. Raychowdhury,
throughput,” in Proc. ISSCC 2014, Feb. 2014,
“Integrated power management in IoT devices
pp. 350–351.
under wide dynamic ranges of operation,” in Proc.
[111] S. Maity, D. Das, and S. Sen, “Wearable health
2015 52nd ACM/EDAC/IEEE Design Autom. Conf.,
monitoring using capacitive voltage-mode human
Jun. 2015, pp. 1–6.
body communication,” in Proc. 2017 39th Ann. Int.
[124] S. B. Nasir, S. Gangopadhyay, and A. Raychowdhury,
Conf. IEEE Eng. Med. Biol. Soc., 2017.
“A 0.13µm fully digital low-dropout regulator with
[112] S. Maity et al., “Secure human-internet using dynamic
adaptive control and reduced dynamic stability for
human body communication,” in Proc. 2017 IEEE/ACM
ultra-wide dynamic range,” in Proc. ISSCC 2015,
Int. Symp. Low Power Electron. Design, 2017, pp. 1–6.
Feb. 2015, pp. 1–3.
[113] S. Maity et al., “Characterization and classification
[125] S. B. Nasir, S. Gangopadhyay, and A. Raychowdhury,
of human body channel as a function of excitation
“All-digital low-dropout regulator with adaptive control
and termination modalities,” arXiv preprint
and reduced dynamic stability for digital load circuits,”
arXiv:1805.02492, 2018.
IEEE Trans. Power Electron., vol. 31, pp. 8293–8302,
[114] S. Sen, “Socialhbc: Social networking and secure
Dec. 2016.
authentication using interference-robust human body
[126] S. B. Nasir, S. Sen, and A. Raychowdhury, “Switched-
communication,” in Proc. 2016 Int. Symp. Low Power
mode-control based hybrid LDO for fine-grain power
Electron. Design, 2016, pp. 34–39.
[115] S. Maity et al., “BioPhysical modeling, management of digital load circuits,” IEEE J. Solid-

characterization and optimization of electro- State Circ., vol. 53, pp. 569–581, Feb. 2018.

quasistatic human body communication,” IEEE Trans. [127] A. Keshavarzi, “Embedded systems and innovative

Biomed. Eng., pp. 1–1, 2018. technologies for IoT applications,” in Proc. Tutorial
[116] S. Maity, K. Mojabe, and S. Sen, “Characterization 62nd Int. ELectron Dev. Meeting, 2016.
of human body forward path loss and variability [128] Z. Ghodsi, S. Garg, and R. Karri, “Optimal
effects in voltage-mode HBC,” IEEE Microw. Wireless checkpointing for secure intermittently-powered
Compon. Lett., vol. 28, no. 3, pp. 266–268, 2018. IoT devices,” in Proc. 2017 IEEE/ACM Int. Conf.
[117] S. Maity et al., “A sub-nw wake-up receiver for human Comput.-Aided Design, Nov. 2017, pp. 376–383.
body communication,” in Proc. 2018 IEEE Biomedical [129] N. Cao et al., “Self-optimizing IoT wireless video
Circ. Syst. Conf., 2018, pp. 1–4. sensor node with in situ data analytics and context-
[118] T. Musah et al., “A 4-32 Gb/s bidirectional link with driven energy-aware real-time adaptation,” IEEE Trans.
3-tap ffe/6-tap dfe and collaborative cdr in 22 nm Circ. and Syst. I, vol. 64, pp. 2470–2480, Sep. 2017.
cmos,” IEEE J. Solid-State Circ., vol. 49, no. 12, [130] N. Cao et al., “Smart sensing for HVAC control:
pp. 3079–3090, 2014. Collaborative intelligence in optical and IR cameras,”

March/April 2019
37
Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 SI: Survey

IEEE Trans. Ind. Electron., vol. 65, pp. 9785–9794, Internet of Things,” IEEE Sensors J., vol. 13,
Dec. 2018. pp. 3677–3684, Oct. 2013.
[131] J. King and A. I. Awad, “A distributed security [146] Z. Li et al., “Research on PKI-like protocol for the Internet
mechanism for resource-constrained IoT devices,” of Things,” in Proc. 2013 Fifth Int. Conf. Meas. Technol.
Informatica 40, vol. 40, pp. 133–143, 2016. Mechatron. Autom., Jan. 2013, pp. 915–918.
[132] L. Atzori, A. Iera, and G. Morabito, “The Internet of [147] E. Hammer-Lahav, The OAuth 1.0 Protocol.
Things: A survey,” Comput. Netw., vol. 54, Accessed: December 12, 2018. [Online]. Available:
pp. 2787–2805, Oct. 2010. https://tools.ietf.org/html/rfc5849
[133] O. Saeed, “Towards Internet of Things: Survey and [148] OAuth 2.0. [Online]. Accessed: December 8, 2018.
future vision,” Int. J. Comput. Netw., 2013. Available: https://oauth.net/2/
[134] M. Yun and B. Yuxin, “Research on the architecture [149] S. Cirani, G. Ferrari, and L. Veltri, “Enforcing security
and key technology of Internet of Things (IoT) applied mechanisms in the IP-based Internet of Things: An
on smart grid,” in Proc. 2010 Int. Conf. Adv. Energy algorithmic overview,” Algorithms, vol. 6, no. 2,
Eng., Jun. 2010, pp. 69–72. pp. 197–226, 2013.
[135] CISCO, The Internet of Things Reference Model. [150] L. Kulseng et al., “Lightweight mutual authentication
Accessed: December 5, 2018. [Online]. Available: and ownership transfer for RFID systems,” in Proc.
https://www.cisco.com/c/dam/global/en_ph/assets/ IEEE INFOCOM, Mar. 2010, pp. 1–5.
ciscoconnect/pdf/bigdata/jim_green_cisco_connect.pdf. [151] G. Zhao et al., “A novel mutual authentication
[136] A. Mosenia and N. K. Jha, “A comprehensive study of scheme for Internet of Things,” in Proc. Int. Conf.
security of Internet-of-Things,” IEEE Trans. Emerg. Modeling, Identif. Contr., Jun. 2011, pp. 563–566.
Topics Comput., vol. 5, pp. 586–602, Oct. 2017. [152] T. Eisenbarth et al., “A survey of lightweight-
[137] M. Burhan et al., “IoT elements, layered architectures cryptography implementations,” IEEE Des. Test,
and security issues: A comprehensive survey,” vol. 24, pp. 522–533, Nov. 2007.
Sensors, vol. 18, no. 9, 2018. [153] S. Ravi et al., “Security in embedded systems: Design
[138] H. Suo, J. Wan, C. Zou, and J. Liu, “Security in the challenges,” ACM Trans. Embed. Comput. Syst.,
Internet of Things: A review,” in Proc. 2012 Int. Conf. vol. 3, pp. 461–491, Aug. 2004.
Comput. Sci. Electron. Eng., vol. 3, pp. 648–651, [154] S. Babar et al., “Proposed embedded security
Mar. 2012. framework for Internet of Things (IoT),” in Proc. 2011
[139] M. V. Bharathi et al., “Node capture attack in wireless 2nd Int. Conf. Wireless VITAE, Feb. 2011, pp. 1–5.
sensor network: A survey,” in Proc. 2012 IEEE Int. [155] A. Sanzgiri and D. Dasgupta, “Classification of insider
Conf. Comput. Intell. Comput. Res., Dec. 2012, threat detection techniques,” in Proc. 11th Ann. Cyber
pp. 1–3. Inf. Security Res. Conf. ’16, New York, NY, 2016,
[140] D. Puthal et al., “Threats to networking cloud and pp. 25:1–25:4.
edge datacenters in the Internet of Things,” IEEE [156] J. R. C. Nurse et al., “Smart insiders: Exploring the
Cloud Comput., vol. 3, pp. 64–71, May 2016. threat from insiders using the Internet-of-Things,” in
[141] D. Brumley and D. Boneh, “Remote timing attacks Proc. 2015 Int. Workshop Secure Internet of Things,
are practical,” in Proc. 12th Conf. USENIX Security Sep. 2015, pp. 5–14.
Symp. ’03, Berkeley, CA, USENIX Association, 2003, [157] S. Madakam, R. Ramaswamy, and S. K. Tripathi,
vol. 12, pp. 1–1. “Internet of Things (IoT): A literature review,”
[142] M. Conti, N. Dragoni, and V. Lesyk, “A survey of man J. Comput. Commun., 2015.
in the middle attacks,” IEEE Commun. Surveys Tuts., [158] R. Khan et al., “Future internet: The Internet of
vol. 18, pp. 2027–2051, Thirdquarter 2016. Things architecture, possible applications and key
[143] R. H. Weber, “Internet of Things—new security and challenges,” in Proc. 2012 10th Int. Conf. Front. Inf.
privacy challenges,” Comput. Law Security Rev., Technol., Dec. 2012, pp. 257–260.
vol. 26, no. 1, pp. 23–30, 2010. [159] P. Sethi and S. R. Sarangi, “Internet of Things:
[144] B. V. Sundaram et al., “Encryption and hash based Architectures, protocols, and applications,” J. Electr.
security in Internet of Things,” in Proc. 2015 3rd Int. Conf. Comput. Eng., 2017.
Signal Process. Commun. Netw., Mar. 2015, pp. 1–6. [160] Q. M. Ashraf and M. H. Habaebi, “Autonomic
[145] F. Li and P. Xiong, “Practical secure communication schemes for threat mitigation in Internet of Things,”
for integrating wireless sensor networks into the J. Netw. Comput. Appl., vol. 49, pp. 112–127, 2015.

38 IEEE Design&Test

Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
[161] R. Canzanese, M. Kam, and S. Mancoridis, “Toward [174] D. Das et al., “STELLAR: A generic EM side-channel
an automatic, online behavioral malware classification attack protection through ground-up root-cause
system,” in Proc. 2013 IEEE 7th Int. Conf. Self-Adaptive analysis,” in Proc. 2019 IEEE Int. Symp. Hardw.
Self-Organizing Syst., Sep. 2013, pp. 111–120. Oriented Security Trust, May 2019.
[162] L. Bilge and T. Dumitra, “Before we knew it: An [175] D. Das et al., “X-DeepSCA: Cross-device deep
empirical study of zero-day attacks in the real learning side channel attack,” in Proc. 2019 ACM/
world,” in Proc. 2012 ACM Conf. Comput. Commun. IEEE Design Autom. Conf., 2019.
Security’12, New York, NY, 2012, pp. 833–844. [176] C. Herder et al., “Physical unclonable functions and
[163] R. Kaur and M. Singh, “A survey on zero-day polymorphic applications: A tutorial,” Proc. IEEE, Aug. 2014,
worm detection techniques,” IEEE Commun. Surveys vol. 102, pp. 1126–1141.
Tuts., vol. 16, pp. 1520–1549, Third 2014. [177] U. Ruhrmair and M. van Dijk, “PUFs in security
[164] A. Bar-On et al., “Improved key recovery attacks on protocols: Attack models and security evaluations,” in
reduced-round AES with practical data and memory Proc. 2013 IEEE Symp. Security Privacy, May 2013,
complexities.” Cryptology ePrint Archive, Report pp. 286–300.
2018/527, 2018. [178] B. Chatterjee, D. Das, and S. Sen, “RF-PUF: IoT
[165] A. Biryukov et al., “Key recovery attacks of practical security enhancement through authentication of
complexity on AES-256 variants with up to 10 wireless nodes using in situ machine learning,” in
rounds,” in Proc. 29th Ann. Int. Conf. Theory Appl. Proc. 2018 IEEE Int. Symp. Hardw. Oriented Security
Cryptogr. Techn. ’10, Berlin, Heidelberg, Springer- Trust, Apr. 2018, pp. 205–208.
Verlag, 2010, pp. 299–319. [179] G. E. Suh and S. Devadas, “Physical unclonable
[166] I. Thomson, AES-256 Keys Sniffed in Seconds functions for device authentication and secret key
Using e200 Kit a Few Inches Away. Accessed: generation,” in Proc. 2007 44th ACM/IEEE Design
November 11, 2018. [Online]. Available: Autom. Conf., Jun. 2007, pp. 9–14.
https://www.theregister.co.uk/2017/06/23/aes_256_ [180] Z. Paral and S. Devadas, “Reliable and efficient PUF-
cracked_50_seconds_200_kit/ based key generation using pattern matching,” in
[167] C. C. E. Brier and F. Olivier, “Correlation power Proc. 2011 IEEE Int. Symp. Hardw.-Oriented Security
analysis with a leakage model,” Cryptogr. Hardw. Trust, Jun. 2011, pp. 128–133.
Embedded Syst., pp. 16–29, 2004. [181] D. E. Holcomb, W. P. Burleson, and K. Fu, “Power-Up
[168] D. Das et al., “ASNI: Attenuated signature noise SRAM state as an identifying fingerprint and source
injection for low-overhead power side-channel attack of true random numbers,” IEEE Trans. Comput.,
immunity,” IEEE Trans. Circ. Syst. I, vol. 65, vol. 58, pp. 11981210, Sep. 2009.
pp. 3300–3311, Oct. 2018. [182] D. E. Holcomb, W. P. Burleson, and K. Fu, “Initial
[169] D. Das et al., “High efficiency power side-channel SRAM state as a fingerprint and source of true
attack immunity using noise injection in attenuated random numbers for RFID tags,” in Proc. Conf. RFID
signature domain,” in Proc. 2017 IEEE Int. Symp. Security, 2007.
Hardw. Oriented Security Trust, 2017, pp. 62–67. [183] Y. Su, J. Holleman, and B. P. Otis, “A Digital 1.6 pJ/
[170] M. Kar et al., “Reducing power side-channel bit chip identification circuit using process variations,”
information leakage of AES engines using fully IEEE J. Solid-State Circ., vol. 43, pp. 69–77, Jan. 2008.
integrated inductive voltage regulator,” IEEE J. Solid- [184] S. Sutar et al., “D-PUF: An intrinsically reconfigurable
State Circ., vol. 53, pp. 2399–2414, Aug. 2018. DRAM PUF for device authentication and
[171] A. Singh et al., “Improved power/EM side-channel random number generation,” ACM Trans. Embed.
attack resistance of 128-bit AES engines with Comput. Syst., vol. 17, pp. 17:1–17:31, Dec. 2017.
random fast voltage dithering,” IEEE J. Solid-State [185] U. Ruhrmair et al., “Modeling attacks on physical
Circ., vol. 54, pp. 569–583, Feb. 2019. unclonable functions,” in Proc. 17th ACM Conf.
[172] T. D. Cnudde, M. Ender, and A. Moradi, “Hardware Comput. Commun. Security ’10, New York, NY, 2010,
masking, revisited,” Trans. Cryptogr. Hardw. pp. 237–249.
Embedded Syst., vol. 2018, pp. 123–148, 2018. [186] C. Zhou, K. K. Parhi, and C. H. Kim, “Secure and
[173] T.-H. Le et al., “A proposition for correlation power reliable XOR Arbiter PUF design: An experimental
analysis enhancement,” in Proc. 8th Int. Conf. study based on 1 trillion challenge response pair
Cryptogr. Hardw. Embedded Syst. ’06, Berlin, measurements,” in Proc. 54th Ann. Design Autom.
Heidelberg, Springer-Verlag, 2006, pp. 174–186. Conf. ’17, New York, NY, 2017, pp. 10:1–10:6.

March/April 2019
39
Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.
 SI: Survey

[187] B. Chatterjee et al., “RF-PUF: Enhancing IoT security [201] C. F. Camerer, T.-H. Ho, and J.-K. Chong, “A cognitive
through authentication of wireless nodes using in situ hierarchy model of games*,” Quart. J. Econ., vol. 119,
machine learning,” IEEE Internet of Things J., 2018. pp. 861–898, Aug. 2004.
[188] M. A. Alsheikh et al., “Machine learning in wireless [202] B. W. Rogers, T. R. Palfrey, and C. F. Camerer,
sensor networks: Algorithms, strategies, and “Heterogeneous quantal response equilibrium and
applications,” IEEE Commun. Surveys Tuts., vol. 16, cognitive hierarchies,” J. Econ. Theory, vol. 144,
pp. 1996–2018, Fourthquarter 2014. no. 4, pp. 1440–1467, 2009.
[189] R. Masiero et al., “Data acquisition through joint
compressive sensing and principal component Baibhab Chatterjee is currently pursuing a PhD
analysis,” in Proc. GLOBECOM 2009—2009 IEEE degree with the School of Electrical Eng­ineering, Pur-
Global Telecommun. Conf., Nov. 2009, pp. 1–6. due University, West Lafayette, IN. His research inter-
[190] T. M. Cover, “Hypothesis testing with finite statistics,” ests include low-power analog, RF, and mixed-signal
Ann. Math. Statist., vol. 40, pp. 828–835, Jun. 1969.
circuit design for secure biomedical/IoT applications.
Chatterjee has a BTech in electronics and communi-
[191] T. Park and W. Saad, “Learning with finite memory for
cation engineering from National Institute of Technol-
machine type communication,” ArXiv. 1610.01723,
ogy Durgapur, Durgapur, India (2011), an MTech in
vol. abs/1610.01723, 2016.
electrical engineering from the IIT Bombay, Mumbai,
[192] K. Drakopoulos, A. E. Ozdaglar, and J. N. Tsitsiklis, India (2015). He is a Student Member of the IEEE.
“On learning with finite memory,” ArXiv. 1209.1122,
vol. abs/1209.1122, 2012. Ningyuan Cao is currently pursuing a PhD degree
[193] R. S. Sutton and A. G. Barto, Introduction to in electrical engineering with the Georgia Institute
Reinforcement Learning. Cambridge, MA: MIT Press, of Technology, Atlanta, GA. His research interests
1998, 1st ed. include low-power machine-learning ASIC design,
[194] L. P. Kaelbling, M. L. Littma, and A. W. Moore, wireless sensor design and power management,
“Reinforcement learning: A survey,” J. Artif. Intell. and energy harvesting circuit design. Cao has a BS
Res., vol. 4, 1996. in electrical engineering from Shanghai Jiaotong
[195] D. Athukoralage et al., “Regret based learning for UAV University, Shanghai, China (2013), an MS in electrical
assisted LTE-U/WiFi public safety networks,” in Proc.
engineering from Columbia University, New York, NY
(2015). He is a Student Member of the IEEE.
2016 IEEE Global Commun. Conf., Dec. 2016, pp. 1–7.
[196] Z. Liu and I. Elhanany, “RL-MAC: A QoS-aware
reinforcement learning based MAC protocol for
Arijit Raychowdhury is an Associate Professor
with the School of Electrical and Computer Engineer-
wireless sensor networks,” in Proc. 2006 IEEE Int.
ing, Georgia Institute of Technology, Atlanta, GA.
Conf. Netw. Sensing Contr., Apr. 2006,
His research interests include low-power digital and
pp. 768–773.
mixed-signal circuit design, device–circuit interac-
[197] A. Amravati et al., “A 55nm time-domain mixed- tions, and novel computing models and hardware
signal neuromorphic accelerator with stochastic realizations. Raychowdhury has a PhD in electrical
synapses and embedded reinforcement learning for and computer engineering from Purdue University,
autonomous micro-robots,” in Proc. 2018 IEEE Int. West Lafayette, IN (2007). He is a Senior Member of
Solid-State Circ. Conf., Feb. 2018, pp. 124–126. the IEEE.
[198] M. Camelo and J. F. a. S. Latre, “A scalable parallel
Q-learning algorithm for resource constrained Shreyas Sen is an Assistant Professor with the
decentralized computing environments,” in Proc. School of Electrical and Computer Engineering,
Purdue University, West Lafayette, IN. His research
2016 2nd Workshop Machine Learn. HPC Environ,
interests include mixed-signal circuits/systems for
Nov. 2016, pp. 27–35.
Human Body Communication, IoT, Biomedical, and
[199] A. M. Printista, M. L. Errecalde, and C. I. Montoya,
Hardware Security. Sen has a PhD in electrical and
“A parallel implementation of Q-learning based
computer engineering from Georgia Tech, Atlanta,
on communication with cache,” in J. Comput. Sci. GA (2011). He is a Senior Member of the IEEE.
Techn., 2002.
[200] R. M. Kretchmar, “Reinforcement learning algorithms  Direct questions and comments about this article
for homogenous multi-agent systems,” in Workshop to Baibhab Chatterjee, Purdue University, West
Agent Swarm Program., 2003. Lafayette, IN 47907, USA; [email protected].

40 IEEE Design&Test

Authorized licensed use limited to: Tampere University. Downloaded on December 30,2022 at 23:15:54 UTC from IEEE Xplore. Restrictions apply.