See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/380548812

Wireless Network Analytics for the New Era of Spectrum Patrolling and
Monitoring

Article in IEEE Wireless Communications · August 2024

DOI: 10.1109/MWC.012.2300374

CITATIONS READS

2 45

2 authors:

Enrico Testi Andrea Giorgetti

University of Bologna University of Bologna
27 PUBLICATIONS 160 CITATIONS 169 PUBLICATIONS 4,003 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Andrea Giorgetti on 15 July 2024.

The user has requested enhancement of the downloaded file.

This article has been accepted for inclusion in a future issue of this magazine. Content is final as presented, with the exception of pagination.
ACCEPTED FROM OPEN CALL

Wireless Network Analytics for the New Era

of Spectrum Patrolling and Monitoring
Enrico Testi and Andrea Giorgetti

Abstract spectrum sensing, localization, network orchestra-

The importance of networks in modern-day tion, and optimization are only a few examples of
society is experiencing rapid and extensive how this promising field will underpin next-gener-
growth, thanks to remarkable technological ation networks.
advancements. Next-generation wireless networks Thanks to these remarkable advances, 6G net-
will confront a scenario where densely intercon- works will have all it takes to become the nervous
nected devices demand high data rates, at least system of our society.
99.999 percent reliability, and extremely low Unfortunately, only limited efforts have been
latency for transmissions. In this context, securing made to ensure the resiliency of this complex
the radio spectrum becomes a crucial task. While nervous system. This aspect is even more exac-
adopting artificial intelligence-based technologies erbated by the upcoming revolution of AI, which
can make networks more intelligent and self-gov- will make our devices intelligent and efficient, as
erning, it also exposes them to new vulnerabilities. already pointed out, but way more vulnerable [5].
Unfortunately, while the networks evolve at a dra- These vulnerabilities can give rise to various
matic pace and the spectrum becomes increasing- forms of unauthorized uses, including scenarios like
ly crowded, the way we monitor its usage appears lower-tier devices encroaching on spectrum allocat-
inadequate and sporadic. Therefore, the need for ed for higher-tier nodes, unapproved devices uti-
methods and procedures for large-scale autono- lizing licensed spectrum through software defined
mous spectrum monitoring seems inevitable to radios, jammers causing disruptions, and denial of
face future challenging scenarios. In this article, service attacks, among the others. For instance,
we present our perspective on a set of tools suit- it is commonly assumed that jammers employ
able for efficient spectrum monitoring, which we attack strategies that remain constant over time;
refer to as wireless network analytics. in contrast, if we suppose that the attackers are
also empowered, for example, with AI capabilities,
Introduction they may potentially adjust their attack strategy in
In the future, with the advent of 6G communi- response to the changing dynamics of the environ-
cation systems, there will be an overwhelming ment and the legitimate communications, becom-
demand for radio spectrum resources by billions ing hard to detect. For this reason, it is expected
of wireless devices such as sensors, connected that, along with 5G and 6G technologies, we will
vehicles, and smart infrastructures [1]. A paradigm witness the dawn of new systems that, using the
shift in wireless radio technology will be necessary same AI tools to achieve unprecedented communi-
to cater to this massive connectivity. Presumably, cations capabilities, will pose severe threats to our
6G will continue to benefit from the technologi- communication-based infrastructures [6].
cal advancements of 5G, but new strategies will As a result, ensuring real-time monitoring of
undoubtedly be needed to make the next step radio spectrum usage to comply with regulations
change. Researchers are already investing sub- becomes of utmost importance. In this regard,
stantial effort in pushing technology to its limits large-scale spectrum patrolling is becoming a
to achieve denser networks, wider bandwidth, critical element for detecting improper spectrum
lower latency, and increased reliability [2]. To sat- usage in a society heavily reliant on communica-
isfy such requirements, devices with intelligent tions technology. With this perspective, gaining
adaptive learning and decision-making capabilities a deeper understanding of how networks utilize
will be key elements [2, 3]. Furthermore, future the wireless medium can significantly contribute
6G devices are envisioned to possess sophisticat- to the development of more effective spectrum
ed spectrum awareness tools based on learning monitoring strategies [7].
and inference methods to optimize the utilization The main contributions of this article may be
of available resources. In 6G networks, artificial summarized as follows:
intelligence (AI) will realize its full potential, pro- • We propose a vision of the challenges and
viding unparalleled autonomous decision-making opportunities regarding the next generation of
and self-regulation capabilities to heterogeneous spectrum monitoring systems in the era of AI.
mobile radio systems, as envisioned over two In particular, we want to outline a framework
decades ago [4]. AI-based methodologies for for spectrum patrolling that leveraging AI and

Digital Object Identifier: 10.1109/MWC.012.2300374 The authors are with the University of Bologna, Italy.

1 1536-1284/24/$25.00
Authorized licensed use limited to: Universita © Downloaded
degli Studi di Bologna. 2024 IEEE on July 15,2024 at 12:10:40
IEEE UTC
Wireless from IEEE Xplore.
Communications Restrictions
• Accepted apply.
for Publication
 Radio spectrum mon-
itoring is a complex
process that aims to
obtain a high-level
logical representation
of the radio frequency
environment, its users,
and their behavior.

FIGURE 1. A depiction of a 6G wireless network disrupted by an intelligent jamming device. A UAV, acting as spectrum
patrol, monitors the RF environment and applies advanced analytics extraction techniques to gain valuable insights from
the data, aiming to obtain as much information as possible about the situation and eventually identify the jammer. The
extracted analytics are then forwarded to the authority in charge so that it can secure the environment.

statistical signal processing can be used for effec- with almost 100% reliability, the monitoring of our
tive monitoring, including classification and local- shared resource, the spectrum, is still reliant on
ization of threats, interference mitigation, and regulators using costly and energy-intensive lab-
detection of violations of spectrum access rules. grade spectrum analyzers. This approach makes
• We define the fundamental ingredients for the spectrum monitoring inefficient, slow, outdated,
composition of a spectrum management tool- sporadic, and unsuitable for scalability. Converse-
kit. In this context, AI plays an ambivalent role: ly, to bolster the security of forthcoming wireless
on the one hand, it will be pervasive in future systems, there is a pressing need for ongoing and
devices and networks, making it necessary to comprehensive spectrum patrolling, that is vital to
monitor the spectrum more attentively and rig- enable rapid responses and ensure comprehen-
orously; on the other hand, it will be a valuable sive coverage across time, frequency, and space
tool to amplify spectrum patrolling capabilities. dimensions [8, 9].
• Numerical results illustrate how it is possible To start with a concrete example, Fig. 1 depicts
to extract wireless network analytics after per- a wireless network scenario in which a spectrum
forming blind source separation (BSS) from raw patroller, represented by a UAV, collects the data
measurement acquired by radio-frequency (RF) about the RF scene and extracts a set of wireless
sensors with remarkable accuracy. network analytics. Such analytics are then forward-
ed to the authority that, according to the spec-
What Is Spectrum Monitoring trum utilization rules defined by the regulator (e.g.,
Radio spectrum monitoring is a complex process European Telecommunications Standards Institute
that aims to obtain a high-level logical representa- (ETSI) or Federal Communications Commission
tion of the radio frequency environment, its users, (FCC)), proceeds to identify anomalies or malicious
and their behavior. It involves three main phases. users (e.g., the jammer) that threaten the network.
Initially, a collection of data in the frequency, time, Spectrum patrolling with UAVs has the potential
and spatial domains through spectrum sensing and to revolutionize the way we manage and protect
patrolling mechanisms is necessary. Then, the mul- our wireless communication infrastructure. Drones
tidimensional data are processed to extract a set of equipped with specialized sensors and software
logical features (or analytics) that give a detailed can fly over large areas, quickly identifying and
representation of the spectrum usage and its users. locating any unauthorized use of the RF spectrum.
Finally, the extracted features are analyzed to iden- This is especially important in large urban areas
tify behaviors or situations for which an action where traditional spectrum patrolling techniques
is required, for example, the intervention of the can be time-consuming and labor-intensive.
authority to interrupt a jammer or reorganize a Another advantage of using UAVs for spectrum
wireless network for more efficient spectrum reuse. patrolling is their ability to access hard-to-reach
areas, that is, flying over mountains, forests, and
Spectrum Sensing and Patrolling other remote zones.
The monitoring process begins with data acqui- The patroller can either be a dedicated device
sition through spectrum sensing techniques. It is, or an adhoc designed network (or fleet) provided
therefore, essential to ponder which actors should by the authority on charge (Fig. 1). Alternatively,
perform this primary task and how they should do patrolling can be performed via crowdsourcing in
it. Indeed, despite the demand for communication which the mobile devices periodically sense the

IEEEAuthorized licensed use limited

Wireless Communications to: Universita
• Accepted degli Studi di Bologna. Downloaded on July 15,2024 at 12:10:40 UTC from IEEE Xplore. Restrictions apply.
for Publication 2
With the widespread spectrum and send condensed data to mobile toring and extracting information about the RF sce-
use of smartphones, operators or the authority [8]. The high number nario. The patrol can collect over-the-air received
almost everyone has of intelligent devices, enhanced by AI capabilities, signals and perform an in-depth analysis of the
access to a device will be able to act as a vast low-cost heterogeneous adversarial network without being part of it. The
network of spectrum sensors. This method involves acquired knowledge about the RF environment
capable of detecting enlisting the help of the general public to identify might help, for example, protect own communica-
and collecting informa- and report any unauthorized use of the RF spec- tions, predict potential interference, and infer hos-
tions about RF signals. trum and has the potential to increase the effective- tile users’ behavior from their network dynamics.
By tapping into this ness and efficiency of spectrum monitoring. Industrial Security: In an industrial environ-
collective knowledge, Spectrum patrolling through crowdsourcing ment, a group of machines operating in an assem-
benefits a large pool of contributors. With the bly line, communicating, and synchronizing their
spectrum patrolling widespread use of smartphones, almost every- tasks through wireless connections, can be sus-
through crowdsourc- one has access to a device capable of detecting ceptible to jamming attacks. A malicious transmit-
ing can cover a larger and collecting informations about RF signals. By ter could disrupt inter-machine communications,
area than traditional tapping into this collective knowledge, spectrum leading to production process slowdowns or halts,
monitoring methods, patrolling through crowdsourcing can cover a resulting in significant inconveniences and sub-
larger area than traditional monitoring methods, stantial financial losses, as well as dangerous situa-
also providing more also providing more detailed data. tions for the workers. To address this vulnerability,
detailed data. Another advantage of crowdsourcing is the deploying a spectrum patrol in the production
potential for real-time reporting. The public can environment becomes essential. The spectrum
report any unauthorized RF signals as soon as they patroller can continuously monitor the wireless
sense them. This allows for a quicker response spectrum, detecting any signs of jamming activity.
and faster mitigation of any potential interference. By swiftly identifying and localizing the jammer,
In some cases, this can even lead to the identi- the spectrum patrol can take appropriate mea-
fication of interference sources that would have sures to secure the communications and ensure
otherwise gone undetected. the smooth operation of the assembly line.
Finally, crowdsourcing can increase public
engagement in the protection of our wireless Challenges in Spectrum Monitoring and Patrolling
communication systems. By allowing the public to While the advantages of spectrum patrolling and
participate in spectrum patrolling, it raises aware- monitoring are evident, addressing numerous chal-
ness about the importance of spectrum manage- lenges remains pivotal. Monitoring extensive por-
ment and encourages citizens to take an active tions of the radio spectrum necessitates specialized
role in safeguarding our critical infrastructure. equipment which can be heavy and cumbersome
Indeed, the adoption of the large-scale crowd- for patrollers. Effectively monitoring vast areas
sourcing paradigm has the potential to revolu- demands strategic deployment of radio sensors
tionize spectrum patrolling, making this tool a and the creation of optimized UAV flight trajec-
valuable and impactful resource [8]. tories. Recent research has explored the UAV tra-
jectory design problem using AI, yet a standard
All-Round Network Awareness: Practical Examples trajectory design and optimization approach for
To enhance the physical layer security of next-gen- UAV swarms for cooperative spectrum monitoring
eration wireless networks, it seems essential, if not is lacking. The information extractable from the
mandatory, that the knowledge of the RF scenario analyzed wireless network plays a crucial role in
is shared across all the actors involved. Indeed, such deploying, orchestrating, and designing the trajec-
knowledge paves the way for a new set of features tories of patrol elements. Consequently, stringent
that take advantage of the all-round awareness of constraints exist on the maximum latency permit-
the networks. Specifically, let us outline a set of sce- ted for feature extraction. Addressing this chal-
narios in which the implementation of a spectrum lenge necessitates the development of innovative
patroller significantly enhances networks’ security. AI algorithms and advanced signal processing tech-
Self-Regulating Networks: A wireless network niques. Moreover, preserving privacy imposes addi-
composed of primary users (PUs) may want to tional constraints on the types of data accessible to
determine whether another network is legitimate- patrols, complicating the extraction of information
ly or illegitimately utilizing the same spectrum. In from wireless networks. This underscores the need
this scenario, the wireless nodes have the capabil- to develop novel, highly reliable algorithms tailored
ity to schedule regular sensing periods to monitor to extract specific data analytics from nonsensitive
the RF medium [10]. Once the primary network datasets, contributing to the advancement of spec-
gathers the RF spectrum samples, it can detect trum monitoring technology.
an unknown adversarial network and perform an
analysis to extract key information about it. Then, Tools for Spectrum Monitoring
the primary network can make informed choices Collecting non-trivial features of the wireless net-
concerning spectrum usage, optimize its oper- work at different levels, ranging from the physical
ations in response to the adversarial network’s layer to the application layer, will play a pivot-
behavior, or report communication violations to al role in detecting anomalies like jammers and
spectrum regulators. In this context, dedicating unauthorized spectrum utilizers. Moreover, it will
time to spectrum sensing seems to decrease the enable the optimization of communications and
primary network’s throughput, yet providing valu- the maximization of spectrum reuse, leading to
able information about spectrum usage which more efficient and robust wireless networks. In
may counterbalance the sensing overhead. this section, we aim to highlight a series of tools
Defense: In a defense scenario, a network of that will empower the patrol to effectively man-
spectrum patrollers can be deployed in a hostile or age and safeguard a next-generation wireless net-
dangerous environment with the purpose of moni- work and its users.

Authorized licensed use limited to: Universita degli Studi di Bologna. Downloaded on July 15,2024 at 12:10:40 UTC from IEEE Xplore. Restrictions apply.
3 IEEE Wireless Communications • Accepted for Publication
 Spectrum Patrol

Blind Source Wireless Network Analytics

Detection - Topology Inference
Authority
& - Transmitters Localization
Separation - Traffic Classification

6G Network Nodes

Malicious User

FIGURE 2. An illustration of the proposed framework. The spectrum patrol gathers over-the-air traffic profiles produced by
the nodes of the 6G wireless network. Subsequently, source detection and BSS are performed to separate the received
signals. The final stage is the extraction of network analytics such as the topology of the observed network, the location of
active transmitters, and the type of application-level traffic generated.

System Setup and Common Assumptions Source Counting and Blind Source Separation
As an example, Fig. 2 shows a spectrum patrol Considering that the patrol will reasonably be com-
composed of a set of RF sensors monitoring posed of multiple devices equipped with antenna
a wireless environment and processing all the arrays, it is expected that a step toward radio envi-
observations to extract network analytics. The ronment monitoring could be identifying a set of
same scenario is illustrated in Fig. 3, where the unknown transmitters and extracting their transmit-
patrollers are UAVs, which survey the scenario ted power profiles. Hence, the first problem relies
by sensing and transmitting all the acquired data on estimating the number of active devices in the
to a fusion center, in charge of merging and pro- observed environment. This task can be accom-
cessing all the observations. The fusion center plished through model order selection using mini-
extracts the wireless network analytics and gets mum description length (MDL) criteria.
a detailed, comprehensive view of the observed The diagram in Fig. 2 represents the complete
RF environment. The listed analytics are the processing sequence. First, each RF sensor collects
main ones studied at the moment, and we are a mixture of over-the-air received powers from all
sure that others will emerge in the near future. the nodes of the observed network. Subsequently,
These features can be extracted without interact- all the gathered traffic profiles are transmitted to
ing with the observed network and the sensors. a fusion center, which can be either one of the
The processing can be accomplished without devices composing the patrol or the authority in
demodulating the received signals, employing charge. The received mixtures are processed to
methods such as a simple energy detector (ED) effectively separate and distinguish the traffic pro-
[11]. This approach does not require collabo- files generated by each network node.1 This recon-
ration between the network and the patrol and struction phase can lay the foundation for a more
preserves the users’ privacy. Conversely, if the effective analytics extraction procedure.
network is collaborative, these characteristics The identification and the separation of the
can be combined with the information made power profiles start from BSS, leveraging on spatial
available by the network itself to obtain more filtering, matrix factorization, tensor decomposition,
effective monitoring. The simulation results illus- and independent component analysis (ICA), to men-
trated in the following subsections are obtained tion a few state-of-the-art techniques, followed by
from the raw measurement of a set RF sensors statistical signal processing methods and ML [11].
deployed in the same area of an IEEE 802.11s The objective is to reconstruct the temporal traffic
ad-hoc network, operating at f 0 = 2.412 GHz. profiles as if they were measured at each transmit-
The propagation scenario is characterized by ting node of the network. Indeed, due to the wire-
omnidirectional antennas at the nodes and the less medium’s nature, the patrol sensors observe a
sensors, path-loss, log-normal shadowing, and mixture of signals emitted by the nodes. To extract
thermal noise. The simulations have been carried and distinguish the individual signals generated by
out for different shadowing intensities, simulat- each node, an unmixing operation is necessary. 1 The separation of signals
ing various urban scenarios. The number of RF The blind source separation algorithm used in would provide more informa-
sensors is greater than or equal to the number this work is based on the combination of princi- tion, but extraction of tempo-
ral power profiles has been
of network nodes, such that the problem of sep- pal component analysis (PCA) and Fast-ICA tech- shown to provide enough
arating the signals emitted by the nodes is over- niques, and is the starting point for the extraction information, for example, for
determined. of the following wireless network analytics. network topology inference.

IEEEAuthorized licensed use limited

Wireless Communications to: Universita
• Accepted degli Studi di Bologna. Downloaded on July 15,2024 at 12:10:40 UTC from IEEE Xplore. Restrictions apply.
for Publication 4
 Patrol unit
pNN
FA pNN
D pGC
FA pGC
D pTE
FA pTE
D
0.8

0.6

p D , p FA
0.4
(a)
0.2

0
2 3 4 5
Network node Shadowing intensity (dB)

20
LSstd

Localization error (m)

LSavg
15
MLstd
MLavg
10

(b)
5

0
1 1.4 1.8 2.2 2.6
Nodes/sensors ratio

Classification accuracy
Spectrum 0.8
Monitoring
0.6 PCA
KPCA
(c)
SVM
0.4 NN

Authority 10 20 30 40 50 60 70 80 90 100
Observation window (ms)

FIGURE 3. An example of a spectrum patrolling scenario where a wireless network is monitored by patrol units (i.e., UAVs)
equipped with RF sensors. The patrol units collect the over-the-air signals transmitted by the nodes and send them to the
fusion center. The fusion center extracts the analytics and gets a detailed network description. Such information can be
forwarded to the authority and used to secure the network from malicious users (i.e., jammers) and optimize spectrum
usage: a) Topology inference: performance of a set of topology inference algorithms, in terms of detection (pD) and false-
alarm (pFA) probabilities of the directed links of pairs of nodes of the monitored network; b) Nodes localization: results of LS
and MLE-based localization performances in terms of localization error varying the ratio between the number of sensors
and the number of network nodes; c) Traffic classification: performance comparison for different ML classifiers, namely
SVMs, NNs, PCA and KPCA, in the recognition of different user-level applications varying the observation time window.

Topology Inference lem to find a causal relationship between two time

The topology of a wireless network can be esti- series; the goal is to infer if there is a causal rela-
mated by accessing the information of the net- tionship between the transmitted power profiles of
work routing protocol. While this approach can a node i and that of node j.
be adopted in some scenarios, the presence of The first widely known definition is Granger
malicious users or the network’s inability to com- causality (GC), which models the time series as
municate with the patroller may require that topol- linear L-th order autoregressive (AR) processes. In
ogy is inferred from outside the network. If this particular, considering a pair of power profiles, xi
task is already challenging for a wired network, it and xj, emitted by node i and node j, respectively,
becomes even more complex for a wireless net- two hypotheses (models) can be formulated [11]
work due to factors like interference, path-loss, ' '
shadowing, fading, and the hidden terminal prob-
H! : 𝑥𝑥",$ = % 𝛽𝛽% 𝑥𝑥",$–% + % γ% 𝑥𝑥),$–% + 𝜀𝜀$
lem. Additionally, while node connectivity might
be deduced from their proximity, identifying %(!
'
%(!
which nodes are actively communicating requires
the analysis of their activity patterns. If we con- H* : 𝑥𝑥",$ = % 𝛿𝛿% 𝑥𝑥",$–% + 𝜔𝜔$
sider a network of N nodes, its logical topology (1) %(!
is represented by a directed graph and its asso- where {bl}Ll=1,
and {dl}Ll=1 are the regression {gl}Ll=1,
ciated adjacency matrix A  {0, 1} NN where coefficients, and ek, wk, are samples of independent
ai,j = 1 if information flows from node i to j and additive white Gaussian noise. The model that cor-
0 otherwise. The literature offers several method- responds to hypothesis H1 considers the possibil-
ologies to infer the topology of a network solely ity of a causal relationship between the two time
by observing the power profiles transmitted (and series, while the null hypothesis H0 excludes the
received) by its nodes. A promising approach is to contribution of the past values of xi in the predic-
exploit the concept of causality to make a logical tion of xj . This means that if xi Granger causes xj
association between power profiles emitted by the prediction error of the model that corresponds
network nodes. More precisely, if there is a data to hypothesis H1 is less than the one of the model
flow from node i to node j, their transmission pro- that corresponds to hypothesis H0. On the other
files are expected to exhibit a causal relationship, hand, if xi has no causal influence on xj the errors
for example, an ACK packet sent by node j can are approximately equal. A F-test can be used to
be the effect of the transmission of node i, that is, decide for one of the two hypotheses.
the cause. A second tool for detecting causality is transfer
Causality can be defined in several ways entropy (TE), a specific type of directed informa-
depending on the context. For brevity, we report tion which, unlike GC, is model-independent and
the two most used metrics that better fit the prob- measures the flow of information between two

Authorized licensed use limited to: Universita degli Studi di Bologna. Downloaded on July 15,2024 at 12:10:40 UTC from IEEE Xplore. Restrictions apply.
5 IEEE Wireless Communications • Accepted for Publication
random processes. Considering two time series Fig. 3(b) shows the results of the non-collabo- Since some devices
xi and xj , modeled as random processes, the TE rative localization procedure proposed in [14], can be malicious or
from i to j can be expressed as [12, 13] where a set of RF sensors perform a received sig- use a technology that
nal strength (RSS)-based localization of the nodes does not include a
𝑝𝑝(𝑥𝑥#,% |𝐱𝐱!– , 𝐱𝐱#– ) of an unknown wireless network through BSS and
TE!→# = I"𝑥𝑥#,% ; 𝐱𝐱!– , 𝐱𝐱#– ' = 𝔼𝔼 )log ' 1 least squares (LS) or maximum likelihood esti- localization procedure
𝑝𝑝(𝑥𝑥#,% |𝐱𝐱#– ) (e.g., Wi-Fi, 2G), the
mation (MLE). The performance indicator is the
(2) localization error, evaluated for different values patrol must manage
where x–i and x–j denote the past samples of xi and of the ratio between the density of RF sensors both collaborative
xj up to time instant k, respectively, and p(·|·) is and network nodes. As evident from the results, and non-collaborative
the conditional probability density function (p.d.f.). when this ratio increases, the error decreases,
The well-known GC is based on a relatively sim- indicating that a higher number of sensors posi- localization to form a
ple test grounded on AR models, an assumption tively impacts the localization performance. The comprehensive picture
that does not always hold. On the contrary, TE results demonstrated that the combination of BSS of the RF scene.
is model-free but requires the estimation of mul- and MLE have satisfactory performance in realistic
tidimensional p.d.f.s, known to be data-hungry channels with noise and shadowing, making the
and computationally demanding. In the context of blind localization of multiple unknown simultane-
topology inference, alternative solutions that rely ous transmitters possible.
on neural network (NN) have been found prac-
tical and computationally appealing after proper Traffic Classification
training [11]. However, causality detection is still The importance of identifying resource usage and
an open problem, and we expect to see interest- coordinating security operations, such as firewall-
ing developments in the coming years. ing and filtering out undesirable traffic, cannot be
Figure 3a illustrates an example of possible overstated in large-scale network infrastructures.
output for topology inference and a way to mea- Understanding the type of traffic generated by the
sure its performance as described in [11], where users of a wireless network is crucial for identifying
pD and pFA indicate the probability of detection malevolent transmitters. These activities encompass
and false alarm of the directed link from two the entire ISO-OSI stack, ranging from the physi-
nodes of the monitored wireless network, cal- cal layer to the application level. For instance, at
culated over a set of Monte Carlo trials. Figure one end, traffic profiles generated by a jammer can
3a shows a comparison between GC, TE, and be detected by identifying irregular transmission
a NN-based causality detection algorithm. Such patterns; on the other one, more in-depth knowl-
results, accounting for packet collisions and edge of how a network uses the wireless medium
realistic channel impairments, such as noise and may contribute to the development of much more
shadowing, revealed that wireless network topol- effective spectrum sharing strategies.
ogy inference of unknown networks is possible Although traffic classification in wired networks
and that the accuracy in detecting information has been extensively studied, few papers address
flows between nodes, that is, reconstructing this issue in wireless systems where the amount of
the adjacency matrix (hence network topology) accessible information is limited. In recent times,
can be quite good, especially for the proposed novel ML-based techniques for traffic classification
AI-based solution. In fact, the NN outperforms in non-collaborative wireless networks employing
the traditional solutions in a low shadowing affordable RF sensors have emerged [15]. These
regime, performing topology detection with approaches leverage machine learning algorithms
lower complexity [11]. to effectively classify the nature of network traf-
fic, even in scenarios where nodes do not actively
Transmitters Localization cooperate or exchange information, showcasing
An essential insight that a patroller can find while the potential of utilizing cost-effective RF sensing
monitoring a network is the spatial position of its for advanced network analysis.
nodes. This analytical data enables the authori- The approach developed in this work can classi-
ty to pinpoint potential sources of interference. fy different types of user-level applications, namely
By combining location information with other video streaming, web browsing, and chat.
extracted insights, the authority gains a compre- Figure 3c depicts a possible output of the clas-
hensive view that aids in effective interference sification stage and shows a performance compar-
mitigation and network security. ison for different classification techniques, such as
Since some devices can be malicious or use a support vector machines (SVMs), neural networks
technology that does not include a localization pro- (NNs), PCA and kernel principal component
cedure (e.g., Wi-Fi, 2G), the patrol must manage analysis (KPCA), that assess their ability to clas-
both collaborative and non-collaborative localization sify traffic patterns. The classifiers’ performance
to form a comprehensive picture of the RF scene. is evaluated in terms of accuracy, defined as the
Most of the works of the past decade assume fraction of correct classifications, as a function
that the targets can exchange information with of the acquisition time window. As a result, the
the patroller, actively contributing to the local- NN outperforms the other classifiers achieving
ization process. While this approach can be remarkable performance in the presence of prop-
exploited, considering that localization of users is agation impairments and with a short observation
a matter of fact in 5G, and even more so in 6G window (just 30 ms). Even in this context, while
networks, we should also envisage the localiza- simple linear PCA-based classifiers require longer
tion of non-collaborative devices either because observation time to distinguish between classes,
the monitored network does not share this type the NN provides the best performance thanks to
of data or because of the presence of malicious its capability of extracting features from heteroge-
users, which are not willing to be localized [14]. neous data and finding non-linear boundaries.

IEEEAuthorized licensed use limited

Wireless Communications to: Universita
• Accepted degli Studi di Bologna. Downloaded on July 15,2024 at 12:10:40 UTC from IEEE Xplore. Restrictions apply.
for Publication 6
In our framework, System Design Guidelines Acknowledgment
spectrum awareness
can leverage patrollers, This section aims to provide readers with a con- This work was partially supported by the Europe-
which, in agreement cise set of guidelines for designing a spectrum an Union under the Italian National Recovery and
patrolling system, incorporating considerations Resilience Plan (NRRP) of NextGenerationEU, part-
with the authority (and related to the challenges outlined above. The first nership on “Telecommunications of the Future”
spectrum regulators), problem to tackle is selecting the adequate num- (PE00000001 – program “RESTART”). The authors
will extract valuable ber and types of patrollers, which could be static, are with the Department of Electrical, Electronic,
features of the observed moving on the ground, or flying, based on the and Information Engineering “Guglielmo Marconi”
networks in multiple intended monitoring approach. Key factors influ- (DEI), CNIT, University of Bologna, Italy.
encing these choices include the scenario size,
domains (frequency, the specific spectrum portions to be monitored, References
time, space, etc.). [1] IBM Institute for Business Value, “Internet of Threats: Secur-
the number and type of network analytics to be ing the Internet of Things for Industrial and Utility Compa-
extracted, device costs, and patrollers’ hardware nies,” 2020; available: https://www.ibm.com/downloads/
size and weight. For instance, a UAV performing cas/ZJRRVRKW.
wideband spectrum monitoring needs to carry [2] G. Gui et al., “6G: Opening New Horizons for Integration of
Comfort, Security, and Intelligence,” IEEE Wirel. Commun.,
weighty equipment, which impacts its energy con- vol. 27, no. 5, Mar. 2020, pp. 126–32.
sumption and consequently influences the design [3] M. S. J. Solaija, H. Salman, and H. Arslan, “Towards a Unified
of its trajectory. Then, deployment and trajecto- Framework for Physical Layer Security in 5G and Beyond
ries of patrollers must be designed to jointly max- Networks,” IEEE Open J. Vehicular Technology, vol. 3, June
2022, pp. 321–43.
imize the performance of signal detection and [4] J. Mitola and G. Maguire, “Cognitive Radio: Making Soft-
separation and network feature extraction algo- ware Radios More Personal,” IEEE Pers. Commun., vol. 6, no.
rithms. The next critical step is the selection or 4, 1999, pp. 13–18.
development of signal detection and separation [5] C. Jiang et al., “Machine learning paradigms for next-genera-
tion wireless networks,” IEEE Wirel. Commun., vol. 24, no. 2,
algorithms. These algorithms play a crucial role in Dec. 2017, pp. 98–105.
extracting information about the nodes within the [6] A. Krayani et al., “Self-Learning Bayesian Generative Models
wireless network under observation, depending for Jammer Detection in Cognitive-UAV-Radios,” Proc. IEEE
on the type of patrollers in use. When the number Global Commun. Conf., Taipei, Taiwan, Dec. 2020, pp. 1–7.
[7] T. J. O’Shea, J. Corgan, and T. C. Clancy, “Convolutional
of patrollers is fewer than the devices to be mon- Radio Modulation Recognition Networks,” Engineering
itored (i.e., underdetermined case), signal sepa- Applications of Neural Networks, Springer International Pub-
ration becomes remarkably complex. Currently, lishing, 2016, pp. 213–26.
there are no algorithms tailored for the effective [8] A. Chakraborty et al., “Spectrum Patrolling With Crowd-
sourced Spectrum Sensors,” Proc. IEEE Conf. Computer
separation of radio signals in the underdeter- Commun., Honolulu, USA, Apr. 2018, pp. 1682–90.
mined case. Finally, the choice of network ana- [9] H. M. Furqan et al., “Wireless Communication, Sensing, and
lytics extraction algorithms hinges upon several REM: A Security Perspective,” IEEE Open J. Commun. Soci-
crucial factors, including the specific network fea- ety, vol. 2, Jan. 2021, pp. 287–321.
[10] L. Arcangeloni, E. Testi, and A. Giorgetti, “Detection of
tures desired, latency constraints, and extraction Jamming Attacks via Source Separation and Causal Infer-
accuracy. Additionally, it is imperative to consider ence,” IEEE Trans. Commun., vol. 71, no. 8, May 2023, pp.
privacy preservation when identifying appropriate 4793–4806.
algorithms for extracting network attributes. [11] E. Testi and A. Giorgetti, “Blind Wireless Network Topology
Inference,” IEEE Trans. Commun., vol. 69, no. 2, Feb. 2021,

Conclusions pp. 1109–20.

[12] J. Massey, “Causality, Feedback and Directed Information,”
This article proposed a vision of the challenges Proc. Int’l. Symp. Inf. Theory Applic., Honolulu, Hawaii, USA,
and opportunities regarding the next generation 1990, pp. 303–05.
[13] T. Schreiber, “Measuring Information Transfer,” Phys. Rev.
of spectrum monitoring systems in the era of AI. Lett., vol. 85, July 2000, pp. 461–64.
While the pervasive use of AI will make 6G net- [14] E. Testi and A. Giorgetti, “RSS-Based Localization of Mul-
works more efficient, their empowered capabili- tiple Radio Transmitters via Blind Source Separation,” IEEE
ties will be more vulnerable than today’s systems. Commun. Lett., vol. 26, no. 03, 2021, pp. 532–36.
[15] E. Testi et al., “Machine Learning for User Traffic Classifi-
In this radical change of networks’ paradigm, cation in Wireless Systems,” Proc. Europ. Sig. Proc. Conf.,
spectrum awareness gains importance, and new Rome, Italy, Sep. 2018, pp. 2040–44.
approaches to monitoring the spectrum on a
large scale will be unavoidable. Biographies
In our framework, spectrum awareness can Enrico Testi [M] received the M.S. degree in electronics and
telecommunications engineering for energy (magna cum laude)
leverage patrollers, which, in agreement with the and the Ph.D. degree in Electronics, Telecommunications, and
authority (and spectrum regulators), will extract valu- Information Technologies engineering from the University of
able features of the observed networks in multiple Bologna, Italy, in 2018 and 2022, respectively. He is currently
domains (frequency, time, space, etc.). Such attri- junior assistant professor at the Department of Electrical, Elec-
tronic, and Information Engineering “Guglielmo Marconi” of
butes, for example, the number and the position the University of Bologna. His research interests include artificial
of the nodes, the application level traffic generated intelligence (AI) techniques for next-generation wireless net-
by them, and the logical topology of the networks, works, massive MIMO, and LEO satellite communications.
can be used as tools to detect malicious users (e.g.,
Andrea Giorgetti [SM] received the Dr. Ing. degree (summa
jammers), manage the spectrum reuse paradigms, cum laude) in electronic engineering and the Ph.D. degree in
coordinate interference, and help cross-network electronic and computer engineering from the University of
orchestration. In light of the considerations and Bologna, Italy, in 1999 and 2003, respectively. He is currently
motivations addressed in this article, as well as the an Associate Professor with the Department of Electrical, Elec-
tronic, and Information Engineering, University of Bologna. His
numerical results illustrated, we believe that spec- research interests include statistical signal processing, integrated
trum patrolling and monitoring will be a fundamen- sensing and communications, localization, and cognitive radio.
tal part of the future wireless ecosystem.

Authorized licensed use limited to: Universita degli Studi di Bologna. Downloaded on July 15,2024 at 12:10:40 UTC from IEEE Xplore. Restrictions apply.
7 IEEE Wireless Communications • Accepted for Publication
View publication stats