Full Compressed Sensing For Privacy Preserving Data Processing Matteo Testa PDF All Chapters
Full Compressed Sensing For Privacy Preserving Data Processing Matteo Testa PDF All Chapters
com
https://textbookfull.com/product/compressed-
sensing-for-privacy-preserving-data-processing-
matteo-testa/
https://textbookfull.com/product/compressed-sensing-for-engineers-
first-edition-majumdar-angshul/
textbookfull.com
https://textbookfull.com/product/secure-and-privacy-preserving-data-
communication-in-internet-of-things-1st-edition-liehuang-zhu/
textbookfull.com
https://textbookfull.com/product/processing-of-remote-sensing-data-
first-edition-hazel-b-girard/
textbookfull.com
https://textbookfull.com/product/neck-rological-1st-edition-elizabeth-
dunlap/
textbookfull.com
Popular Sovereignty in Early Modern Constitutional Thought
Daniel Lee
https://textbookfull.com/product/popular-sovereignty-in-early-modern-
constitutional-thought-daniel-lee/
textbookfull.com
Touch The Complete Series 1st Edition Cara Dee Dee Cara
https://textbookfull.com/product/touch-the-complete-series-1st-
edition-cara-dee-dee-cara/
textbookfull.com
https://textbookfull.com/product/an-introduction-to-advanced-
mathematics-mirroslav-yotov/
textbookfull.com
https://textbookfull.com/product/behavior-space-play-pleasure-and-
discovery-as-a-model-for-business-value-1st-edition-manu/
textbookfull.com
Epidemics and the Modern World Hammond
https://textbookfull.com/product/epidemics-and-the-modern-world-
hammond/
textbookfull.com
SPRINGER BRIEFS IN ELEC TRIC AL AND
COMPUTER ENGINEERING SIGNAL PROCESSING
Compressed
Sensing for
Privacy-Preserving
Data Processing
123
SpringerBriefs in Electrical and Computer
Engineering
Signal Processing
Series editors
Woon-Seng Gan, School of Electrical and Electronic Engineering,
Nanyang Technological University, Singapore, Singapore
C.-C. Jay Kuo, Los Angeles, CA, USA
Thomas Fang Zheng, Tsinghua University, Beijing, China
Mauro Barni, Università degli Studi di Siena, Siena, Italy
More information about this series at http://www.springer.com/series/11560
Matteo Testa Diego Valsesia
•
Compressed Sensing
for Privacy-Preserving Data
Processing
123
Matteo Testa Tiziano Bianchi
Department of Electronics Department of Electronics
and Telecommunications and Telecommunications
Politecnico di Torino Politecnico di Torino
Turin, Italy Turin, Italy
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2019
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made. The publisher remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Preface
v
vi Preface
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Compressed Sensing and Security . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1 Compressed Sensing as a Cryptosystem . . . . . . . . . . . . . . . . . . . . 11
2.1.1 Security Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.2 Attack Scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.2 Signal Embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3 Compressed Sensing as a Cryptosystem . . . . . . . . . . . . . . . . . . . . . . 25
3.1 Statistical Properties of Measurements . . . . . . . . . . . . . . . . . . . . . 26
3.2 Gaussian Sensing Matrices and Asymptotic Behavior . . . . . . . . . . 30
3.2.1 Model Definition and Security Metrics . . . . . . . . . . . . . . . 30
3.2.2 Energy Obfuscation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.3 Upper Bound Validation . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2.4 Asymptotic Behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.3 Arbitrary Sensing Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3.1 Model Definition and Security Metrics . . . . . . . . . . . . . . . 40
3.3.2 Generic Unstructured Sensing Matrices . . . . . . . . . . . . . . . 44
3.3.3 Circulant Sensing Matrices . . . . . . . . . . . . . . . . . . . . . . . . 52
3.3.4 Upper Bound Validation . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.4 Practical Sensing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.4.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.4.2 Sensing Matrix Classes . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.4.3 Sensing Matrix Generation . . . . . . . . . . . . . . . . . . . . . . . . 68
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
vii
viii Contents
4 Privacy-Preserving Embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.1 User Authentication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.1.2 System Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.1.3 Security Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.2 Bounded-Distance Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.2.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.2.2 Universal Embeddings . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.2.3 Private Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Chapter 1
Introduction
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2019 1
M. Testa et al., Compressed Sensing for Privacy-Preserving Data Processing,
SpringerBriefs in Signal Processing, https://doi.org/10.1007/978-981-13-2279-2_1
Visit https://textbookfull.com
now to explore a rich
collection of eBooks, textbook
and enjoy exciting offers!
2 1 Introduction
highly valuable information. Moreover, it appears more and more evident that the
large mass of IoT devices may not have sufficient capabilities for deploying con-
ventional cryptographic solutions. Many of these devices are battery operated and
are often left unattended, with very limited maintenance, limiting both their power
consumption and computational capabilities [27].
Among the solutions able to meet the stringent requirements of IoT devices, com-
pressed sensing (CS) can be considered as a very promising option. CS is a mature
technology enabling simultaneous signal acquisition and compression, based on rep-
resenting a signal with a small number of highly incoherent linear projections. The
possibility of implementing CS through hardware acquisition (see, e.g., [8]) reduces
the number of required sensing elements, limiting the overall power consumption
[9, 10, 20]. This latter aspect makes the CS framework an excellent candidate for
low-energy devices [16].
At the same time, it has been recognized that the inherent randomness in the
CS acquisition process provides some secrecy guarantees. For example, in [21] the
authors show that CS is computationally secure as long as the sensing matrix is
used only once. Additional security properties of CS were later studied in [3–7, 18],
showing that in the best possible scenario CS measurements leak only the energy of
the sensed signal, and a framework for securing IoT devices through CS has been
recently proposed in [17].
Providing a lightweight encryption layer for low-power devices is not the only
security feature of CS. If one is not concerned with signal recovery, CS measurements
can be modeled as signal embeddings projecting a signal in a low dimensional space
in which distances are approximately preserved [12]. Due to the properties of embed-
dings, some authors suggest that CS can implicitly provide a privacy preserving layer
enabling simple processing tasks [1], like privacy preserving data mining [15], sparse
regression [34], or achieving differential privacy [14]. Several applications relying
on privacy properties of CS measurements have been recently proposed, including
outsourcing image data to the cloud for privacy-preserving data mining and image
retrieval [29, 31, 32], generating a robust image hash [23], providing biometric tem-
plate protection [2, 19, 24, 25], and implementing physical unclonable functions [11,
22, 30]. The ability to perform basic signal processing operations on confidential data
is also beneficial for IoT devices, for example to detect anomalies [26].
In this book, we will try to present the vast amount of literature on the security of
CS under a unifying framework. The scenario we are referring to is that exemplified
in Fig. 1.1. We assume that a number of low-power sensor nodes are transmitting
privacy-sensitive data to a cloud service for enabling several information processing
tasks. Different users can interact with the cloud to obtain the results of the different
tasks. In the above scenario, we identify two weaknesses that can be targeted by
adversaries. First, the communication channel between the sensor node and the cloud
can be attacked by an eavesdropper trying to get access to sensitive information.
Second, the cloud can include some non-trusted entities that observe the collected
information and use this knowledge for malicious purposes.
In the first case, security is achieved using the CS framework as a lightweight
cryptosystems providing some level of secrecy. While here we focus on the
1 Introduction 3
different sensing matrix constructions. For sensing matrices made of Gaussian i.i.d.
entries, we have the highest secrecy guarantees, where only the energy of the signal
can be revealed. This particular case is analyzed by introducing a secrecy metric that
depends on the ability to estimate the signal energy by an adversary who observes
only the signal measurements. The secrecy achievable by generic sensing matrices is
analyzed by introducing a distinguishability metric inspired by the standard statisti-
cal secrecy definition used in cryptography. Results are provided for matrices made
of i.i.d entries with generic distributions and circulant matrices. At the end of the
chapter, we discuss several issues connected with the practical implementation of a
CS cryptosystem, including sensing matrix generation and quantization of sensing
matrix entries.
In Chap. 4, we illustrate main results on privacy-preserving embeddings. Here,
security properties of embeddings are analyzed by considering two possible scenarios
for their use. In the first case, a client submits a query containing sensitive information
to a server, which should respond to the query without gaining access to the private
information. This is discussed describing an authentication system in which a client
submit an embedding of a physical characteristic of a device, and a verification server
is able to match the embedding without revealing the actual physical characteristic.
Interestingly, in this case the security properties of the embedding permit to combine
it with existing biometric template mechanisms, enhancing the security of the system.
In the second case, a large amount of sensitive data is stored in the cloud and a user
should be able to make specific queries to the cloud without gaining access to the
data. Here, we describe a universal embedding that preserves distances only locally.
If data are stored in the cloud using this embedding, a user is able to retrieve data
close to the query, but the complete geometry of the dataset remains hidden by the
embedding and data cannot be recovered.
Finally, Chap. 5 summarizes the main results discussed in the book, providing
some discussion on open issues and promising avenues for future research on this
topic.
References
1. Abdulghani, A., Rodriguez-Villegas, E.: Compressive sensing: from compressing while sam-
pling to compressing and securing while sampling. In: 2010 Annual International Conference
of the IEEE Engineering in Medicine and Biology Society (EMBC), pp. 1127–1130 (2010)
2. Anzaku, E.T., Sohn, H., Ro, Y.M.: Multi-factor authentication using fingerprints and
user-specific random projection. In: 2010 12th International Asia-Pacific Web Conference
(APWEB), pp. 415–418 (2010)
3. Bianchi, T., Bioglio, V., Magli, E.: On the security of random linear measurements. In: 2014
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP’14), pp.
3992–3996 (2014)
4. Bianchi, T., Bioglio, V., Magli, E.: Analysis of one-time random projections for privacy pre-
serving compressed sensing. IEEE Trans. Inf. Forensics Secur. 11(2), 313–327 (2016)
References 5
5. Cambareri, V., Haboba, J., Pareschi, F., Rovatti, H., Setti, G., Wong, K.W.: A two-class informa-
tion concealing system based on compressed sensing. In: 2013 IEEE International Symposium
on Circuits and Systems (ISCAS), pp. 1356–1359 (2013)
6. Cambareri, V., Mangia, M., Pareschi, F., Rovatti, R., Setti, G.: Low-complexity multiclass
encryption by compressed sensing. IEEE Trans. Signal Process. 63(9), 2183–2195 (2015)
7. Cambareri, V., Mangia, M., Pareschi, F., Rovatti, R., Setti, G.: On known-plaintext attacks to
a compressed sensing-based encryption: a quantitative analysis. IEEE Trans. Inf. Forensics
Secur. 10(10), 2182–2195 (2015)
8. Duarte, M.F., Davenport, M.A., Takhar, D., Laska, J.N., Sun, T., Kelly, K.F., Baraniuk, R.G.:
Single-pixel imaging via compressive sampling. IEEE Signal Process. Mag. 25(2), 83–91
(2008)
9. Gangopadhyay, D., Allstot, E.G., Dixon, A.M., Natarajan, K., Gupta, S., Allstot, D.J.: Com-
pressed sensing analog front-end for bio-sensor applications. IEEE J. Solid-State Circuits 49(2),
426–438 (2014)
10. Herman, M.A., Strohmer, T.: High-resolution radar via compressed sensing. IEEE Trans. Signal
Process. 57(6), 2275–2284 (2009)
11. Holotyak, T., Voloshynovskiy, S., Koval, O., Beekhof, F.: Fast physical object identification
based on unclonable features and soft fingerprinting. In: 2011 IEEE International Conference
on Acoustics, Speech and Signal Processing (ICASSP), pp. 1713–1716 (2011)
12. Johnson, W.B., Lindenstrauss, J.: Extensions of Lipschitz mappings into a Hilbert space. Con-
temp. Math. 26 (1984)
13. Katz, J., Lindell, Y.: Introduction to Modern Cryptography. Chapman & Hall/CRC Cryptogra-
phy and Network Security Series. Chapman & Hall/CRC, London (2007)
14. Li, Y.D., Zhang, Z., Winslett, M., Yang, Y.: Compressive mechanism: utilizing sparse represen-
tation in differential privacy. In: Proceedings of the 10th Annual ACM Workshop on Privacy
in the Electronic Society, WPES’11, pp. 177–182. ACM, New York (2011)
15. Liu, K., Kargupta, H., Ryan, J.: Random projection-based multiplicative data perturbation
for privacy preserving distributed data mining. IEEE Trans. Knowl. Data Eng. 18(1), 92–106
(2006)
16. Mangia, M., Marchioni, A., Pareschi, F., Rovatti, R., Setti, G.: Administering quality-energy
trade-off in IOT sensing applications by means of adapted compressed sensing. IEEE J. Emerg.
Sel. Top. Circuits Syst. 1–1 (2018)
17. Mangia, M., Pareschi, F., Rovatti, R., Setti, G.: Low-cost security of IOT sensor nodes with
rakeness-based compressed sensing: statistical and known-plaintext attacks. IEEE Trans. Inf.
Forensics Secur. 13(2), 327–340 (2018)
18. Orsdemir, A., Altun, H., Sharma, G., Bocko, M.: On the security and robustness of encryption
via compressed sensing. In: 2008 IEEE Military Communications Conference (MILCOM
2008), pp. 1–7 (2008)
19. Pillai, J.K., Patel, V.M., Chellappa, R., Ratha, N.K.: Secure and robust IRIS recognition using
random projections and sparse representations. IEEE Trans. Pattern Anal. Mach. Intell. 33(9),
1877–1893 (2011)
20. Quinsac, C., Basarab, A., Girault, J.M., Kouamé, D.: Compressed sensing of ultrasound images:
sampling of spatial and frequency domains. In: 2010 IEEE Workshop on Signal Processing
Systems (SIPS), pp. 231–236. IEEE (2010)
21. Rachlin, Y., Baron, D.: The secrecy of compressed sensing measurements. In: 2008 46th Annual
Allerton Conference on Communication, Control, and Computing, pp. 813–817. IEEE (2008)
22. Shariati, S., Jacques, L., Standaert, F.X., Macq, B., Salhi, M.A., Antoine, P.: Randomly driven
fuzzy key extraction of unclonable images. In: 2010 IEEE International Conference on Image
Processing, pp. 4329–4332 (2010)
23. Sun, R., Zeng, W.: Secure and robust image hashing via compressive sensing. Multimed. Tools
Appl. 70(3), 1651–1665 (2014)
24. Teoh, A.B.J., Goh, A., Ngo, D.C.L.: Random multispace quantization as an analytic mechanism
for biohashing of biometric and random identity inputs. IEEE Trans. Pattern Anal. Mach. Intell.
28(12), 1892–1901 (2006)
6 1 Introduction
25. Teoh, A.B.J., Yuang, C.T.: Cancelable biometrics realization with multispace random projec-
tions. IEEE Trans. Syst. Man Cybern. Part B (Cybern.) 37(5), 1096–1106 (2007)
26. Testa, M., Bianchi, T., Magli, E.: Energy obfuscation for compressive encryption and pro-
cessing. In: 2017 IEEE Workshop on Information Forensics and Security (WIFS), pp. 1–6
(2017)
27. Trappe, W., Howard, R., Moore, R.S.: Low-energy security: limits and opportunities in the
internet of things. IEEE Secur. Priv. 13(1), 14–21 (2015)
28. Troncoso-pastoriza, J.R., Perez-Gonzalez, F.: Secure signal processing in the cloud: enabling
technologies for privacy-preserving multimedia cloud processing. IEEE Signal Process. Mag.
30(2), 29–41 (2013)
29. Valsesia, D., Coluccia, G., Bianchi, T., Magli, E.: Large-scale image retrieval based on com-
pressed camera identification. IEEE Trans. Multimed. 17(9), 1439–1449 (2015)
30. Valsesia, D., Coluccia, G., Bianchi, T., Magli, E.: User authentication via PRNU-based physical
unclonable functions. IEEE Trans. Inf. Forensics Secur. 12(8), 1941–1956 (2017)
31. Wang, Q., Zeng, W., Tian, J.: Compressive sensing based secure multiparty privacy preserving
framework for collaborative data-mining and signal processing. In: 2014 IEEE International
Conference on Multimedia and Expo (ICME), pp. 1–6 (2014)
32. Wang, Q., Zeng, W., Tian, J.: A compressive sensing based secure watermark detection and
privacy preserving storage framework. IEEE Trans. Image Process. 23(3), 1317–1328 (2014)
33. Wu, T., Ruland, C.: Authenticated compressive sensing imaging. In: 2017 International Sym-
posium on Networks, Computers and Communications (ISNCC), pp. 1–6 (2017)
34. Zhou, S., Lafferty, J., Wasserman, L.: Compressed and privacy-sensitive sparse regression.
IEEE Trans. Inf. Theory 55(2), 846–866 (2009)
Chapter 2
Compressed Sensing and Security
Abstract In this chapter we briefly review the Compressed Sensing (CS) frame-
work, discussing the acquisition model, the conditions under which the signal can be
recovered, and the main reconstruction algorithms. Then, we show how CS is essen-
tially analogous to a private key cryptosystem if signal acquisition, signal recovery,
and sensing matrix generation are interpreted as encryption, decryption, and key gen-
eration functions respectively. The basic security properties of this CS cryptosystem
under different attack scenarios are discussed according to standard security defi-
nitions. This sets the basis for the identification of the attack scenarios that will be
analyzed more in depth in Chap. 3. In the second part of this chapter, we introduce the
concept of signal embeddings, which can be seen as a generalization of CS measure-
ments. The properties of some of the most common embeddings are briefly reviewed,
followed by a discussion on how embeddings can provide privacy-preserving func-
tionalities in particular settings.
y = x (2.1)
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2019 7
M. Testa et al., Compressed Sensing for Privacy-Preserving Data Processing,
SpringerBriefs in Signal Processing, https://doi.org/10.1007/978-981-13-2279-2_2
8 2 Compressed Sensing and Security
Fig. 2.1 Compressed sensing. A small number of measurements y of a signal x is acquired with a
sensing matrix . The signal x admits a sparse representation under basis
i.e., trying to minimize the 0 pseudonorm, i.e., the number of nonzero entries of
a vector. However, solving the above problem is NP-hard. The breakthrough result
was realizing that a convex relaxation of this optimization problem, using the 1
norm instead of the 0 pseudonorm, is equivalent to the combinatorial optimization
problem above under certain conditions. This results in the following optimization
problem, which can be solved efficiently:
This is also known as the Basis Pursuit problem [17]. An alternative formulation casts
the problem into its unconstrained version, and is typically referred to as Lasso:
By virtue of Lagrange multipliers, the Lasso and Basis Pursuit formulations are
equivalent for a specific choice of λ.
From now on, for the remainder of this book, unless differently specified we will
assume that is the identity matrix for convenience of explanation, i.e. signal x
is k-sparse. Figure 2.2 shows an intuitive way of explaining why minimizing the 1
norm leads to sparse solutions for n = 2 and m = 1. The line represents the space
2 Compressed Sensing and Security 9
of solutions such that y = x. One can grow the 1 ball until it touches the space of
solutions, and see that its shape promotes a sparse solution.
Nevertheless, requirements on the structure of both the sensing matrix and the orig-
inal signal have to be satisfied in order to make the original signal x the unique solu-
tion to the above problem. These requirements can be summarized by the Restricted
Isometry Property (RIP) [15].
x22 ≤ (1 + δk )x22
(1 − δk )x22 ≤
That is, if a matrix satisfies the RIP, then the energy of k-sparse signals is approx-
imately preserved in the compressed domain. The RIP can be used to establish a
guarantee on the reconstruction performance of the Basis Pursuit reconstruction
algorithm as reported in the following theorem.
σk (x)1
x̂ − x2 ≤ C √
k
being σk (x)1 = xk − x1 the 1 norm of the approximation of signal x with a k-
sparse version obtained keeping only the entries with largest magnitude.
Notice that σk (x)1 = 0 when the signal is exactly k-sparse, so perfect reconstruction
can be achieved.
10 2 Compressed Sensing and Security
Therefore, the RIP gives us a design criterion to identify the classes of sensing
matrices which allow a successful recovery. Several constructions of sensing matrices
have been explored in the literature. The most important ones are random matrices
with i.i.d. sub-Gaussian entries, i.e. whose tails decay as fast as the ones of the
Gaussian distribution. Such matrices satisfy the RIP with high probability for m >
ck log(n/k) for some positive constant c [2]. However, they require the generation of
mn random samples and the computation of the full matrix-vector product to obtain
the measurements, which may be expensive. Faster solutions have been studied by
introducing some structure in the sensing matrix. Examples include sparse random
matrices [26], matrices constructed from expander graphs [31], Toeplitz and circulant
matrices [40], block-diagonal random matrices [23].
CS allows to achieve significant undersampling factors, i.e. a number of measure-
ments much smaller than the original dimensionality of the signal. However, when
considering compression, real-valued measurements must be quantized in order to
obtain a representation using a finite rate. It is possible to quantize CS measurements
down to 1 bit [30] while still being able to recover the original signal, provided
that the reconstruction is “consistent”, i.e., the measurements of the reconstructed
signal must fall into the original quantization bins. It is important to notice that the
naive choice of a uniform scalar quantization of the measurements causes CS-based
compression methods to suffer from poor rate-distortion performance. Essentially,
the undersampling performed during CS acquisition behaves like oversampling when
sparsity is taken into account and its performance falls in line with established results
on scalar quantization of oversampled signals [27]. In a nutshell, while increasing
the number of quantization levels achieves an exponential reduction in the distortion
of the reconstructed signal, increasing the number of measurements only provides
a linear reduction in distortion. Recently, novel quantizer designs have been studied
to achieve an exponential reduction in distortion [7, 28, 44].
It is worth noting that the literature on CS has seen the development of a multitude
of reconstruction algorithms, which can be broadly categorized into three classes:
methods based on convex optimization, methods based on greedy algorithms, and
methods based on approximate message passing. The class of methods based on
optimization stems from the original 1 minimization formulation in (2.3) [3, 9, 13,
14, 46]. Interior-point optimization methods typically have high computational com-
plexity resulting in slow or prohibitive running time. To speed up the computation,
iterative and greedy algorithms have been proposed to perform the optimization.
Among the iterative algorithms we mention proximal gradient methods [1, 6, 18,
19, 29, 29, 47]. They are more accurate than greedy algorithms, at the cost of higher
computational complexity. As an example, Algorithm 1 reports the pseudocode for
the Iterative Soft Thresholding algorithm (ISTA) which is based on proximal gradi-
ent descent. Soft thresholding is the proximal mapping operator to the 1 norm and
intuitively promotes sparsity.
2 Compressed Sensing and Security 11
Algorithm 1 ISTA
Input: Sensing matrix Φ, measurements y,
shrinkage operator ηλ [·] = sgn(·) max(| · | − λ, 0)
Set x(0) = 0, iterate
for t = 1 to StopIter do
x(t) ← ηλ [x(t−1) + ΦT (y − Φx(t−1) )]
end for
ri = y − Φ(Φ†S (i) y)
end for
x̂ = Φ†S (i) y
The fact that CS can be efficiently implemented using randomly generated sensing
matrices means that CS measurements are naturally equipped with some secrecy
notions, that allow us to emply CS as an effective cryptosystem. Indeed, because of
Visit https://textbookfull.com
now to explore a rich
collection of eBooks, textbook
and enjoy exciting offers!
12 2 Compressed Sensing and Security
GenK GenK
x
Φ Φ
y x̂
Rec(Φ, y)
its structure, the CS acquisition model lends itself to providing confidentiality, since
in order to recover a signal given its measurements, the knowledge of the sensing
matrix is necessary. Therefore, only those parties who have access to the sensing
matrix should be able to recover the original signal. Starting from this consideration,
if a party does not have access to the sensing matrix but has only access to
the measurements y, is it then possible to prove that the original signal cannot be
recovered and hence that CS can also provide secrecy? The short answer to this
extensively investigated question is yes: the CS acquisition can act as an encryption
function. However, if we consider a complete private key cryptosystem, we still need
to define a suitable decryption function. Interestingly, CS already provides algorithms
which can be used to recover the original signal and thus can be formulated as
decryption functions.
Let us formalize these concepts and reconsider CS under a cryptographic perspec-
tive. A compressive cryptosystem, as depicted in Fig. 2.3, can be defined as follows.
The signal x is the plaintext, the measurements y are the ciphertext, and the sens-
ing matrix is the secret key. The system is composed of the following functional
blocks:
Encryption The encryption e (x) is performed through CS acquisition as defined
in (2.1), i.e., in a compressive cryptosystem, the encryption function e (x) = x
is a simple linear combination of the plaintext with random weights. In our model,
we assume that the ciphertext y = e (x) is received without errors. In practical CS
applications, it is customary to assume that measurements y are contaminated by
some noise due, e.g., to quantization or channel introduced errors. While this may be
an issue concerning the recovery performance of CS, it does not affect the security
of the system. As it will become clear in the following, thanks to the data processing
inequality any degradation of the measurements will not introduce any advantage for
an adversary.
2.1 Compressed Sensing as a Cryptosystem 13
Even though the secret key is not known, the linear acquisition process of CS may
lead to leaks of information through the measurements. The analysis of the informa-
tion leakage, which is covered in Chap. 3, relies on information theoretic tools and
definitions which we briefly recall here. It is important to highlight that the informa-
tion theoretic approach is stronger than the computational one since it characterizes
the amount of information an attacker can have access to. If no sufficient information
is available, then even with unbounded computational capabilities the attack cannot
succeed.
Given the cryptosystem we defined above, different metrics can be used to char-
acterize its security properties. From an information theoretic perspective, a cryp-
tosystem is said to achieve perfect secrecy if
P[y|x] = P[y]
where y denotes the ciphertext and x denotes the plaintext. Namely, the posterior
probability of the ciphertext given the plaintext is independent of the plaintext. This
implies that an attacker cannot be more successful than random guessing the plaintext.
Perfect secrecy is very difficult to achieve in practical cryptosystems, so this
definition is usually relaxed by introducing some additional assumptions. If we allow
the adversary to have a tiny advantage with respect to random guessing, we say that
a cryptosystem achieves statistical secrecy. More formally, statistical secrecy can
be defined by the following game. Let us consider two messages x1 and x2 , and
randomly encrypt one as y. An adversary observes y and decides whether this is the
encryption of x1 or x2 . The encryption is said to provide statistical secrecy if
1
P[success] = + (2.4)
2
where is negligible with respect to the size of the key. In the following, we will
show that several implementations of the CS cryptosystem achieve a weak form of
statistical secrecy, in which does not decrease exponentially with the size of the
key.
The above definitions are quite strong, since they rely only on the statistical prop-
erties of CS measurements, without making any assumption on the resources of the
adversary. However, they usually hold under the ideal assumption that measurements
are continuous and sensing matrices are modeled by continuous probability distribu-
tions. For practical cases in which both measurements and sensing matrices will be
represented using a finite, although very large, set of possible values, another widely
used definition is that of computational secrecy. In this case, a cryptosystem is said
to be computationally secure if the definition in (2.4) holds for any adversary limited
to algorithms that run in time polynomial with respect to the size of the key.
2.1 Compressed Sensing as a Cryptosystem 15
Table 2.1 Summary of security properties of compressive cryptosystems under different attack
scenarios. CS: same sensing matrix is re-used; OTS: sensing matrix re-generated at each encryption.
The table indicates whether the system can provide some notion of secrecy or not. OTS is discussed
in Chap. 3. Legend: S: statistical secrecy, C: computational secrecy
COA KPA CPA CCA
CS C [38] No No No
OTS S, C [5, 10] C [11] C No
Signal representations seek to capture all the signal information in a compact manner
in order to increase a system efficiency in terms of storage or processing require-
ments. The previous section introduced representations where the primary goal was
to reconstruct the original signal from the compact encoding with a low distortion,
or, under some conditions, perfectly. However, this is only one facet of the signal
representation problem. Oftentimes, we are not directly interested in the signal itself,
but rather in the information that can be extracted from it through processing. It is
therefore the case that this inference process can be performed more efficiently if the
representation of the signal seeks to preserve the information relevant to the process-
ing algorithm rather the the entire signal. A classic example of this is information
retrieval, where one has to decide which among many stored signals better matches
a query template according to some problem-dependent criterion, e.g., a distance in
a metric space. Such problem is not concerned with the actual signals but rather the
geometry of the entire set of signals in the form of their pairwise distances. Similarly,
many problems in machine learning such as clustering, anomaly detection, regres-
sion, classifiers rely on the assumption that the information relevant for the solution
of the problem is encoded in the geometry of the signal set and a proper model using
the distances between signals achieves the desired solution. Hence, a smart signal
representation should seek to provide compact codes in a space with a distance func-
tion that is fast to evaluate and that approximately preserves the distances in the
original space. This would allow an advantageous trade-off between computational
efficiency and performance in terms of the problem-dependent metric.
Embeddings are transformations of a set of signals from a high-dimensional metric
space to a low-dimensional space such that the geometry of the set is approximately
Random documents with unrelated
content Scribd suggests to you:
The Project Gutenberg eBook of Doktor Senki
(2. kötet)
This ebook is for the use of anyone anywhere in the United States
and most other parts of the world at no cost and with almost no
restrictions whatsoever. You may copy it, give it away or re-use it
under the terms of the Project Gutenberg License included with this
ebook or online at www.gutenberg.org. If you are not located in the
United States, you will have to check the laws of the country where
you are located before using this eBook.
Language: Hungarian
DOKTOR SENKI
REGÉNY
II. KÖTET