Compressed Sensing for Privacy Preserving Data

Processing Matteo Testa download

https://textbookfull.com/product/compressed-sensing-for-privacy-
preserving-data-processing-matteo-testa/

Download full version ebook from https://textbookfull.com

We believe these products will be a great fit for you. Click
the link to download now, or visit textbookfull.com
to discover even more!

Compressed sensing for engineers First Edition Majumdar

Angshul

https://textbookfull.com/product/compressed-sensing-for-
engineers-first-edition-majumdar-angshul/

Secure and Privacy Preserving Data Communication in

Internet of Things 1st Edition Liehuang Zhu

https://textbookfull.com/product/secure-and-privacy-preserving-
data-communication-in-internet-of-things-1st-edition-liehuang-
zhu/

Processing Of Remote Sensing Data First Edition Hazel

B. Girard

https://textbookfull.com/product/processing-of-remote-sensing-
data-first-edition-hazel-b-girard/

Preserving Privacy Against Side-Channel Leaks: From

Data Publishing to Web Applications 1st Edition Wen
Ming Liu

https://textbookfull.com/product/preserving-privacy-against-side-
channel-leaks-from-data-publishing-to-web-applications-1st-
edition-wen-ming-liu/
CMOS Circuits for Biological Sensing and Processing 1st
Edition Srinjoy Mitra

https://textbookfull.com/product/cmos-circuits-for-biological-
sensing-and-processing-1st-edition-srinjoy-mitra/

Image processing and GIS for remote sensing :

techniques and applications Second Edition Liu

https://textbookfull.com/product/image-processing-and-gis-for-
remote-sensing-techniques-and-applications-second-edition-liu/

Big Data Privacy Preservation for Cyber Physical

Systems Miao Pan

https://textbookfull.com/product/big-data-privacy-preservation-
for-cyber-physical-systems-miao-pan/

Canning Preserving For Dummies Amelia Jeanroy

https://textbookfull.com/product/canning-preserving-for-dummies-
amelia-jeanroy/

Algorithms for Data and Computation Privacy 1st Edition

Alex X. Liu

https://textbookfull.com/product/algorithms-for-data-and-
computation-privacy-1st-edition-alex-x-liu/
SPRINGER BRIEFS IN ELEC TRIC AL AND
COMPUTER ENGINEERING  SIGNAL PROCESSING

Matteo Testa · Diego Valsesia

Tiziano Bianchi · Enrico Magli

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
•

Tiziano Bianchi Enrico Magli

•

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

Diego Valsesia Enrico Magli

Department of Electronics Department of Electronics
and Telecommunications and Telecommunications
Politecnico di Torino Politecnico di Torino
Turin, Italy Turin, Italy

ISSN 2191-8112 ISSN 2191-8120 (electronic)

SpringerBriefs in Electrical and Computer Engineering
ISSN 2196-4076 ISSN 2196-4084 (electronic)
SpringerBriefs in Signal Processing
ISBN 978-981-13-2278-5 ISBN 978-981-13-2279-2 (eBook)
https://doi.org/10.1007/978-981-13-2279-2

Library of Congress Control Number: 2018954021

© 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

Compressed sensing is an established technique for simultaneous signal acquisition

and compression, as well as dimensionality reduction, based on representing a
signal with a small number of random projections. Some applications are concerned
with the recovery of the original signal from the random projections, which is
possible using nonlinear optimization techniques, while other applications are
mostly concerned with embedding the geometry of a set of signals, thus working
directly on the random projections.
The former class of applications includes the emerging field of Internet of
Things, where a multitude of devices acquires and transmits sensitive information.
However, such devices are typically power-constrained or have very simple com-
putational capabilities, while at the same time, being required to ensure the confi-
dentiality of the transmitted messages. Consequently, conventional acquisition and
information protection methods are a poor fit for this scenario. Replacing such
techniques with compressed sensing is an appealing solution, as it could offer a
security layer on top of signal acquisition and compression without additional cost.
However, its use as a cryptosystem requires a careful theoretical analysis to ensure
the soundness of the method. The latter class of applications includes problems in
information retrieval and signal detection, where one is interested in the results to
their query without revealing the sensitive information contained in the signal itself.
Embeddings obtained with random projections are appealing as they can provide a
privacy-preserving functionality without compromising the performance of the
embedding in terms of accuracy or computational efficiency.
This book is a unique contribution which analyzes this exciting and timely topic
by collecting many different contributions, scattered in the literature, as well as
recent advances in the field, in a single resource with a unified notation. The
objective of this book is to provide the reader with a comprehensive survey of the
topic in an accessible format. The reader is guided in exploring the topic by first
establishing a shared knowledge about compressed sensing and how it is used
nowadays. Then, clear models and definitions for its use as a cryptosystem and a
privacy-preserving embedding are laid down, before tackling state-of-the-art results
for both applications. The reader will conclude the book having learned that the

v
vi Preface

current results in terms of security of compressed techniques allow it to be a very

promising solution to many practical problems of interest. We believe that this book
can find a broad audience among researchers, scientists, or engineers with very
diverse backgrounds, having interests in security, cryptography, and privacy in
information retrieval systems. Accompanying software is made available on the
authors’ Web site (www.ipl.polito.it) to reproduce the experiments and techniques
presented in the book. The only background required to the reader is a good
knowledge of linear algebra, probability, and information theory.

Turin, Italy Matteo Testa

July 2018 Diego Valsesia
Tiziano Bianchi
Enrico Magli
Contents

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

Abstract Information processing systems have been revolutionized by recent

advances in several technological areas, like device miniaturization, wireless trans-
mission, network infrastructure. Traditional information sources have been replaced
by a multitude of devices with sensing capabilities, the so-called Internet of Things
(IoT), including smart home devices, cars with autonomous driving functions,
portable medical devices. Meanwhile, single processing and storage units have been
replaced by cloud services, leading to interconnected systems that share and process
huge amount of data. While this provides endless opportunities to tackle different
societal needs, it poses several problems regarding the security and privacy of the
involved data. This chapter introduces the challenges and the techniques discussed
in the literature to address them and serves as an overview of the book.

Information processing systems have been revolutionized by recent

advances in several technological areas, like device miniaturization, wireless trans-
mission, network infrastructure. Traditional information sources have been replaced
by a multitude of devices with sensing capabilities, the so-called Internet of Things
(IoT), including smart home devices, cars with autonomous driving functions,
portable medical devices. Meanwhile, single processing and storage units have been
replaced by cloud services, leading to interconnected systems that share and process
huge amount of data.
While this provides endless opportunities to tackle different societal needs, and
results in obvious benefits for everyone, at the same time, it poses several problems
regarding the security and privacy of the involved data. Entrusting sensitive data
to remote services is prone to privacy issues, since the user usually has little or
no control on the actual servers managing the computations. Many sensing devices
may be equipped with very basic encryption and authentication capabilities, or no
capabilities at all, being vulnerable to cybersecurity attacks.
Protecting information processing system from different security threats has been
an active area of research in recent years. For example, secure signal processing
techniques have been proposed for privacy-preserving cloud processing [28]. Nev-
ertheless, such techniques are often based on costly cryptographic primitives and
complex protocols, which limit their usefulness to scenarios involving critical or

© 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
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

Fig. 1.1 Compressed sensing for privacy preserving data processing

confidentiality of data, it is worth noting that the envisaged CS cryptosystem acts as

a private key cryptosystem, that can be easily converted in an authentication mech-
anism by using message authentication codes [13, 33]. In the second case, secure
processing is enabled by the privacy preserving properties of CS measurements, that
can be considered as secure embeddings of the underlying signals.
The materials we are going to present are organized in the following chapters.
In Chap. 2, we briefly review the CS framework, discussing the acquisition model,
the conditions under which the signal can be recovered, and the main reconstruc-
tion algorithms. Then, we show how CS is essentially analogous to a private key
cryptosystem if signal acquisition, signal recovery, and sensing matrix generation
are interpreted as encryption, decryption, and key generation functions. The basic
security properties of this CS cryptosystem under different attack scenarios are dis-
cussed according to standard security definitions, identifying the scenarios that will
be analyzed more in depth in the following chapter. In the second part of Chap. 2, we
introduce the concept of signal embeddings, which can be seen as a generalization
of CS measurements. The properties of some of the most common embeddings are
briefly reviewed, then we discuss how embeddings can provide privacy-preserving
functionalities in particular settings.
Chapter 3 presents the most relevant results on CS used as a cryptosystem. First, we
analyze the statistical properties of CS measurements, showing that they always con-
vey at least the energy of the sensed signal. Then, we discuss the secrecy achievable by
4 1 Introduction

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.

Compressed Sensing (CS) [21] is nowadays a popular approach used to simulta-

neously acquire and compress signals within a single operation. Acquisition and
compression can be modeled with a linear operation as

y = x (2.1)

where x ∈ Rn ,  ∈ Rm×n , y ∈ Rm and m  n. The vector y is a more compact rep-

resentation of signal x, hence the name “compressed” sensing, and its entries are
referred to as measurements. The number of measurements required to represent a
signal can be smaller than what is dictated by the Nyquist–Shannon sampling theo-
rem as such theorem only presents a sufficient condition. Signal reconstruction from
a set of measurements is possible provided the signal admits a low-dimensional rep-
resentation. Typically, such low-dimensionality assumption is in the form of sparsity
in a transform domain, i.e., when the signal is represented under a suitable basis, only
few coefficients are nonzero. Alternatively, a signal is said to be “compressible” if
the magnitude of its coefficients decays quickly in some domain, e.g., with a power
law. Figure 2.1 shows a graphical depiction of CS.

© 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 

The breakthrough that started the whole field of CS is that reconstruction of a

signal from its measurements is possible by solving a convex optimization problem.
Let us consider a signal x admitting a sparse representation x =  θ , with θθ 0 =
|{θθ i = 0, i ∈ [1, n]}| = k; this signal is said to be k-sparse if  is the identity matrix,
and k-sparse in a transform domain otherwise. Given a vector of measurements y, the
reconstruction of x is carried out by exploiting the knowledge of the sensing matrix  ,
basis  , and the fact that θ is sparse. The most intuitive way to put this knowledge into
the form of an optimization problem leads to the following combinatorial problem:

θ̂θ = arg min θθ 0 s.t. y =  θ (2.2)

θ

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:

θ̂θ = arg min θθ 1 s.t. y =  θ . (2.3)

θ

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:

θ̂θ = arg min θθ 1 + λy −  θ 22 .

θ

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

Fig. 2.2 Signal

reconstruction via 1
minimization

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].

Definition 2.1 (Restricted Isometry Property) A matrix  satisfies the restricted

isometry property (RIP) of order k if there exists a δk ∈ (0, 1) such that

x22 ≤ (1 + δk )x22
(1 − δk )x22 ≤ 

for all k-sparse vectors x.

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.

√ (Theorem 1.1 of [12]) Suppose that  satisfies the RIP of order 2k

Theorem 2.1
with δ2k < 2 − 1, then the solution x̂ to (2.3) obeys

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

Greedy algorithms [41, 43], generally, build up an approximation to the solu-

tion one step at a time by making locally optimal choices at each step. Examples
include Orthogonal Matching Pursuit (OMP) [42], Stagewise OMP (StOMP) [22],
Regularized OMP (ROMP) [36], and CoSaMP [35]. Although greedy algorithms are
extremely fast, they generally require a larger number of measurements that may not
be easy to acquire. Algorithm 2 reports the pseudocode for the OMP algorithm, where
ST  S )−1 ST denotes the pseudoinverse of the sensing matrix restricted to
 †S = (
the columns in set S. OMP builds the support of the sparse signal by selecting one
entry at a time. This entry is chosen as the one that maximizes the inner product
between the corresponding column of the sensing matrix and a residual, initialized
as the measurements.
Algorithm 2 OMP
Input: Sensing matrix Φ, measurements y, sparsity k
Set r(0) = y, x(0) = 0, S (0) = ∅, iterate
for i = 1 to k do
g(i) ← ΦT r(i−1)
(i)
|gj |
j (i) ← arg maxj Φj 2
S (i) ← S (i−1) ∪ j
(i)

ri = y − Φ(Φ†S (i) y)
end for
x̂ = Φ†S (i) y

Finally, the most recent class of reconstruction algorithm is based on approximate

message passing techniques. The most important of these algorithms is Generalized
Approximate Message Passing (GAMP) [39]. GAMP takes a Bayesian estimation
approach to signal recovery and tackles the computational intractability by reducing
a vector-valued estimation problem into a sequence of scalar problems.

2.1 Compressed Sensing as a Cryptosystem

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
12 2 Compressed Sensing and Security

Fig. 2.3 Scheme of a

Key
compressive cryptosystem
K

GenK GenK

x
Φ Φ

y x̂
Rec(Φ, y)

Encryption eΦ (x) Decryption dΦ (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

Decryption The decryption d (y) corresponds to any CS recovery algorithm that

can be used to recover the original signal. The goal of this function is to recover
the plaintext given the knowledge of both the secret key and the ciphertext. It can
be defined as d (y) = Rec(y,  ), where Rec(y,  ) correspond to any suitable CS
recovery algorithm, e.g. those described earlier in this section. As introduced at the
beginning of this chapter, if some conditions on the number of measurements and on
the structure of the sensing matrix are satisfied, then sparse or approximately sparse
signals can be recovered with overwhelming probability. Thus, different existing
recovery algorithms can be cast into the decryption function d (y). Here, it is impor-
tant to highlight that in general the decryption function is able to approximately
invert the result of the encryption, i.e., d (e (x)) = x̂ ∼ x, rather than exactly as it
happens in regular private key cryptosystems.1 While this does not pose major issues
for CS acquisition, since in many scenarios exact recovery of the original signal is
not always needed, it has some effect on the security of the CS cryptosystem, since
even an approximate decryption may be considered a successful attack.
Key Generation The key of the CS cryptosystem, i.e., the sensing matrix , is
generated at both encryption and decryption side using a common key generation
function GenK that relies on a shared secret key K. This solution avoids transmitting
a large number of matrix entries to the receiving party. The easiest solution is to
design a deterministic key generation function, so that, given K, every plaintext
x is encrypted using the same sensing matrix. However, in the following we will
mostly consider a randomized key generation function that gives a different sensing
matrix at each sensing step. This strategy will be named one-time sensing (OTS),
since it is similar in principle to the well-known one-time pad cryptosystem. As
will become clear in the following sections, OTS is essential for guaranteeing the
security of the CS cryptosystem. In this book, we will assume that GenK is an ideal
function that generates sensing matrices exactly according to a desired distribution.
In practice, cryptographic key derivation functions such as SHA-3 (Keccak) [4] can
be employed to generate the entries of the secret key. For the interested reader, this
particular problem has been addressed in more detail in, e.g., [20, 24, 25].
From the above description, it is evident that a CS cryptosystem is designed
according to Kerckhoffs’ principle [33]: both encryption and decryption functions are
public, the security relies only on the secrecy of the sensing matrix  , or equivalently,
the shared secret key K. Nevertheless, the compressive cryptosystem we just defined
has some vulnerabilities which have to be addressed by a careful analysis of its
components. For this reason, we need to introduce appropriate security definitions
and discuss in detail the main assumptions to be satisfied in order to make the
cryptosystem resilient to common attack scenarios.

1 Exact reconstruction is possible if x is k-sparse, as described earlier in this section.

14 2 Compressed Sensing and Security

2.1.1 Security Definitions

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

2.1.2 Attack Scenarios

In modern cryptography attacks against encryption schemes are usually classified

into four types, according to the resources of the adversary:
• Ciphertext-only attack (COA): In this scenario the adversary observes only the
ciphertext y (or multiple ciphertexts) and tries to recover the plaintext;
• Known-plaintext attack (KPA): In this case the adversary obtains pairs of plain-
text/ciphertext encrypted with the same key and tries to decrypt other ciphertexts;
• Chosen-plaintext attack (CPA): Here, the adversary can choose arbitrary plain-
texts and obtain the corresponding ciphertext, with the aim of decrypting other
ciphertexts;
• Chosen-ciphertext attack (CCA): In this last scenario, the adversary can obtain
the plaintext corresponding to arbitrary ciphertexts, excluding the ciphertexts that
are currently under attack.
The linearity of the acquisition process immediately implies that a compressive
cryptosystem cannot be secure under either CPA or KPA when the same secret key
is re-used at different encryptions. In the case of CPA, if the sensing matrix is kept
fixed, by choosing the canonical basis as plaintext, the ciphertext will exactly reveal
the columns of the secret key  . Differently, in the case of KPA, the secret key cannot
be trivially revealed as in the previous scenario; however given n linearly indepen-
dent plaintext-ciphertext pairs it is possible to solve the system of mn unknowns
corresponding to the entries of  . Thus, a KPA attack will also succeed.
Luckily, the OTS strategy can effectively address both attack scenarios. Re-
generating the secret key at each encryption will make recovering the sensing matrix
unfeasible under both CPA and KPA, since both attack strategies rely on different
and subsequent encryptions to be performed with the same key in order to extract
its full content. For this reason, in Chap. 3 we will concentrate our attention on OTS
cryptosystems. Moreover, if we assume that different sensing matrices are indepen-
dently generated, KPA or CPA do not provide any advantage over COA, so all the
results will be derived under the COA scenario.
It is worth noting that a practical key generation function will unavoidably intro-
duce correlations among sensing matrices, which may be exploited under either KPA
or CPA. Although this analysis is out of the scope of this book, the interested reader
can find some preliminary results in [11].
Regarding the security of the CS cryptosystem under the COA scenario, it is
important to notice that a RIP-satisfying sensing matrix, although being required
for signal recovery, is not necessarily the best option for providing secrecy. Being
the RIP an 2 -norm preserving property, it is easy to foresee that the energy of the
plaintext will be leaked through the measurements. For this reason, it is important
to analyze the secrecy related to the linear acquisition model and characterize the
amount of information that is leaked through the measurements given a specific
choice of sensing matrix. This important problem is addressed in Chap. 3, where the
information leakage of different sensing matrices is analyzed.
16 2 Compressed Sensing and Security

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

Lastly, we can immediately verify that a CS cryptosystem cannot be secure under

CCA, even when using the OTS strategy. Given a targeted encryption y, under CCA
the adversary can obtain the decryption x of y = y + n, where n is an arbitrarily
small error vector. Due to the linearity of the acquisition process, x will be close to x,
meaning that the adversary can always obtain a good approximation of the plaintext,
or distinguish encryptions of different plaintexts with a success probability close
to one. The security properties of compressive cryptosystems under different attack
scenarios are summarized in Table 2.1.

2.2 Signal Embeddings

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
2.2 Signal Embeddings 17

preserved (or conveniently distorted, in some cases). Distances in the original metric
space are replaced by distances measured directly in the low-dimensional space.
Thanks to the lower dimensionality and, possibly, a distance function that is faster to
compute, inference problems only requiring geometric information about the signal
set are greatly accelerated and require less storage.
Definition 2.2 (Embedding) A function f : X → Y is a (g, δ, )-embedding of met-
ric space (X , dX ) into metric space (Y, dY ) if, for all u, v ∈ X , it satisfies

(1 − δ)g (dX (u, v)) − ε ≤ dY ( f (u), f (v)) ≤ (1 + δ)g (dX (u, v)) + ε.

The quantities δ and  are multiplicative and additive ambiguities due to the embed-
ding, respectively. Notice that, in general, distances dY in the embedding space may
be distorted with respect to original distances dX due to map g(·).
The most famous result concerning embeddings is the Johnson - Lindenstrauss
(JL) lemma [32]. Johnson and Lindestrauss demonstrated that it is possible to use a
Lipschitz function f to implement an embedding that preserves Euclidean distances
up to a multiplicative distortion.

Lemma 2.1 (Johnson–Lindenstrauss) Let δ ∈ (0, 1). For every  set X of |X | points in
Rn , if m is a positive integer such that m = O δ −2 log(|X |) , there exists a Lipschitz
mapping f : Rn → Rm such that

(1 − δ)u − v22 ≤  f (u) − f (v)22 ≤ (1 + δ)u − v22 (2.5)

for all u, v ∈ X .

It is interesting to notice that the number of measurements m required to make

the embedding work depends only on the desired fidelity on the Euclidean distances
between points, controlled by the distortion parameter δ, and by the number of signals
in the set to be embedded. Crucially, there is no dependence on the dimensionality
n of input space.
The JL lemma posits the existence of an embedding approximately preserving
Euclidean distances. A practical implementation satisfying the Johnson-
Lindenstrauss lemma can be obtained by using a random linear mapping (also called
random projection)
y = f (x) =  x

for x ∈ X ⊂ Rn and y ∈ Y ⊂ Rm and a random matrix  . It is easy to show that if

the entries of  are drawn independently from specific distributions, then pairwise
Euclidean distances between points in X are equal, within a multiplicative distortion
δ, to the distances between their embeddings in Y, with high probability. In particular,
if the entries of  are independent and identically distributed according to a Gaussian,
Rademacher, or uniform distribution then Eq. (2.5) holds with probability greater
than 1 − c1 elog |X |−c2 δ m for some universal constants c1 and c2 . Simple proofs for
2

such result can be found [2]. Typically they are composed of two main steps: (i)
Discovering Diverse Content Through
Random Scribd Documents
This same motor question controls the proposal for increased speed.
The use of a reserve motor would again increase weights; though
not necessarily in proportion to the aggregate engine capacity.
Perhaps something may be accomplished with a gasoline turbine,
when one is developed. In any case, no sudden increase in speeds
seems to be probable; any further lightening of motors must be
undertaken with deliberation and science. If much higher maximum
speeds are attained, there will be an opportunity to vary the speed
to suit the requirements. Then clutches, gears, brakes, and speed-
changing devices of various sorts will become necessary, and the
problem of weights of journal bearings—already no small matter—
will be made still more serious. And with variable speed must
probably come variable sail area—in preference to tilting—so that
the fabric must be reefed on its frame. Certainly two men, it would
seem, will be needed!
Better methods for starting are required. The hydroplane idea
promises much in this respect. With a better understanding and
control of the conditions associated with successful and safe descent
—perhaps with improved appliances therefor—the problem of ascent
will also be partly solved. If such result can be achieved, these
measures of control must be made automatic.
The building of complete aeroplanes to standard designs would be
extremely profitable at present prices, which range from $2500 to
$5000. Perhaps the most profitable part would be in the building of
the motor. The framing and fabric of an ordinary monoplane could
easily be constructed at a cost below $300. The propeller may cost
$50 more. The expense for wires, ropes, etc., is trifling; and unless
special scientific instruments and accessories are required, all of the
rest of the value lies in the motor and its accessories. Within
reasonable limits, present costs of motors vary about with the horse-
power. The amateur designer must therefore be careful to keep
down weight and power unless he proposes to spend money quite
freely.
 The Case of the Dirigible
Not very much is being heard of performances of dirigible balloons
just at present. They have shown themselves to be lacking in
stanchness and effectiveness under reasonable variations of
weather. We must have fabrics that are stronger for their weight and
more impervious. Envelopes must be so built structurally as to resist
deformation at high speeds, without having any greatly increased
weight. A cheap way of preparing pure hydrogen gas is to be
desired.
Most important of all, the balloon must have a higher speed, to
make it truly dirigible. This, with sufficient steering power, will
protect it against the destructive accidents that have terminated so
many balloon careers. Here again arises the whole question of
power in relation to motor weight, though not as formidably as is the
case with the aeroplane. The required higher speeds are possible
now, at the cost merely of careful structural design, reduced radius
of action, and reduced passenger carrying capacity.
Better altitude control will be attained with better fabrics and the use
of plane fin surfaces at high speeds. The employment of a vertically-
acting propeller as a somewhat wasteful but perhaps finally
necessary measure of safety may also be regarded as probable.
 Giraudon’s Wheel Aeroplane

The Orthopter
The aviplane, ornithoptère or orthopter is a flying machine with bird-
like flapping wings, which has received occasional attention from
time to time, as the result of a too blind adherence to Nature’s
analogies. Every mechanical principle is in favor of the screw as
compared with any reciprocating method of propulsion. There have
been few actual examples of this type: a model was exhibited at the
Grand Central Palace in New York in January of this year.
The mechanism of an orthopter would be relatively complex, and the
flapping wings would have to “feather” on their return stroke. The
flapping speed would have to be very high or the surface area very
great. This last requirement would lead to structural difficulties.
Propulsion would not be uniform, unless additional complications
were introduced. The machine would be the most difficult of any
type to balance. The motion of a bird’s wing is extremely
complicated in its details—one that it would be as difficult to imitate
in a mechanical device as it would be for us to obtain the structural
strength of an eagle’s wing in fabric and metal, with anything like
the same extent of surface and limit of weight. According to
Pettigrew, the efficiency of bird and insect flight depends largely
upon the elasticity of the wing. Chatley gives the ratio of area to
weight as varying from fifty (gnat) to one-half (Australian crane)
square feet per pound. The usual ratio in aeroplanes is from one-
third to one-half.
About the only advantages perceptible with the orthopter type of
machine would be, first, the ability “to start from rest without a
preliminary surface glide”; and second, more independence of
irregularity in air currents, since the propulsive force is exerted over
a greater extent than is that of a screw propeller.

The Helicopter
The gyroplane or helicopter was the type of flying machine regarded
by Lord Kelvin as alone likely to survive. It lifts itself by screw
propellers acting vertically. This form was suggested in 1852. When
only a single screw was used, the whole machine rotated about its
vertical axis. It was attempted to offset this by the use of vertical
fin-planes: but these led to instability in the presence of irregular air
currents. One early form had two oppositely-pitched screws driven
by a complete steam engine and boiler plant. One of the Cornu
helicopters had adjustable inclined planes under the two large
vertically propelling screws. The air which slipped past the screws
imposed a pressure on the inclined planes which was utilized to
produce horizontal movement in any desired direction—if the wind
was not too adverse. A gasoline engine was carried in a sort of well
between the screws.
 Bréguet Gyroplane During Construction
(Helicopter type)
The helicopter may be regarded as the limiting type of aeroplane,
the sail area being reduced nearly to zero; the wings becoming mere
fins, the smaller the better. It therefore requires maximum motor
power and is particularly dependent upon the development of an
excessively light motor. It is launched and descends under perfect
control, without regard to horizontal velocity. It has very little
exposed surface and is therefore both easy to steer and independent
of wind conditions. By properly arranging the screws it can be amply
balanced: but it must have a particularly stout and strong frame.
The development of this machine hinges largely on the propeller. It
is not only necessary to develop power (which means force
multiplied by velocity) but actual propulsive vertical force: and this
must exceed or at least equal the whole weight of the machine.
From ten to forty pounds of lifting force per horse-power have been
actually attained: and with motors weighing less than five pounds
there is evidently some margin. The propellers are of special design,
usually with very large blades. Four are commonly used: one, so to
speak, at each “corner” of the machine. The helicopter is absolutely
dependent upon its motors. It cannot descend safely if the power
fails. If it is to do anything but ascend and descend it must have
additional propulsive machinery for producing horizontal movement.
 Composite Types
The aeroplane is thus particularly weak as to stability, launching, and
descending: but it is economical in power because it uses the air to
hold itself up. The dirigible balloon is lacking in power and speed,
but can ascend and descend safely, even if only by wasteful
methods; and it can carry heavy weights, which are impossible with
the structurally fragile aeroplane. The helicopter is wasteful in power,
but is stable and sure in ascending and descending, providing only
that the motor power does not fail.
Why, then, not combine the types? An aeroplane-dirigible would be
open to only one objection: on the ground of stability. The dirigible-
helicopter would have as its only disadvantage a certain
wastefulness of power, while the aeroplane-helicopter would seem to
have no drawback whatever.
All three combinations have been, or are being, tried. An Italian
engineer officer has designed a balloon-aeroplane. The balloon is
greatly flattened, or lens-shaped, and floats on its side, presenting
its edge to the horizon—if inclination be disregarded. With some
inclination, the machine acts like an aeroplane and is partially self-
sustaining at any reasonable velocity.
The use of a vertically-acting screw on a dirigible combines the
features of that type and the helicopter. This arrangement has also
been the subject of design (as in Captain Miller’s flexible balloon) if
not of construction. The combination of helicopter and aeroplane
seems especially promising: the vertical propellers being employed
for starting and descending, as an emergency safety feature and
perhaps for aid in stabilizing. The fact that composite types of flying
machine have been suggested is perhaps, however, an indication
that the ultimate type has not yet been established.
 What is Promised
The flying machine will probably become the vehicle of the explorer.
If Stanley had been able to use a small high-powered dirigible in the
search for Livingstone, the journey would have been one of hours as
compared with months, the food and general comfort of the party
would have been equal in quality to those attainable at home, and
the expense in money and in human life would have been relatively
trifling.

Wellman’s America
(From Wellman’s Aerial Age)
Most readers will remember the fate of Andrée, and the projected
polar expeditions of Wellman in 1907 and 1909. Misfortune
accompanied both attempts; but one has only to read Peary’s story
of the dogged tramp over the Greenland ice blink to realize that
danger and misfortune in no less degree have accompanied other
plans of Arctic pioneering. With proper design and the right men, it
does not seem unreasonable to expect that a hundred flying
machines may soar above Earth’s invisible axial points during the
next dozen years. 3
The report of Count Zeppelin’s Spitzbergen expedition of last year
has just been made public. This was undertaken to ascertain the
adaptability of flying machines for Arctic navigation. Besides speed
and radius of action, the conclusive factors include that of freedom
from such breakdowns as cannot be made good on the road.
For exploration in other regions, the balloon or the aeroplane is sure
to be employed. Rapidity of progress without fatigue or danger will
replace the floundering through swamps, shivering with ague, and
bickering with hostile natives now associated with tropical and other
expeditions. The stereoscopic camera with its scientific adjuncts will
permit of almost automatic map-making, more comprehensive and
accurate than any now attempted in other than the most settled
sections. It is not too much to expect that arrangements will be
perfected for conducting complete topographical surveys without
more than occasional descents. If extremely high altitudes must be
attained—over a mile—the machines will be of special design; but as
far as can now be anticipated, there will be no insurmountable
difficulties. The virgin peaks of Ruwenzori and the Himalayas may
become easily accessible—even to women and children if they desire
it. We may obtain direct evidence as to the contested ascent of Mt.
McKinley. A report has been current that a Blériot monoplane has
been purchased for use in the inspection of construction work for an
oil pipe line across the Persian desert; the aeroplane being regarded
as “more expeditious and effectual” than an automobile.
The flying machine is the only land vehicle which requires no
“permanent way.” Trains must have rails, bicycles and automobiles
must have good roads. Even the pedestrian gets along better on a
path. The ships of the air and the sea demand no improvement of
the fluids in which they float. To carry mails, parcels, persons, and
even light freight—these applications, if made commercially
practicable tomorrow, 4 would surprise no one; their possibility has
already been amply demonstrated. With the dirigible as the
transatlantic liner and the aeroplane as the naphtha launch of the
air, the whole range of applications is commanded. Hangars and
landing stages—the latter perhaps on the roofs of buildings,
revolutionizing our domestic architecture—may spring up as rapidly
as garages have done. And the aeroplane is potentially (with the
exception of the motorcycle) the cheapest of self-propelled vehicles.
Governments have already considered the possibilities of aerial
smuggling. Perhaps our custom-house officers will soon have to
watch a fence instead of a line: to barricade in two dimensions
instead of one. They will need to be provided with United States
Revenue aeroplanes. But how are aerial frontiers to be marked? And
does a nation own the air above it, or is this, like the high seas, “by
natural right, common to all”? Can a flying-machine blockade-runner
above the three-mile height claim extraterritoriality?
The flying machine is no longer the delusion of the “crank,” because
it has developed a great industry. A now antiquated statement put
the capitalization of aeroplane manufactories in France at a million
dollars, and the development expenditure to date at six millions.
There are dozens of builders, in New York City alone, of
monoplanes, biplanes, gliders, and models. A permanent exhibition
of air craft is just being inaugurated. We have now even an
aeronautic “trust,” since the million-dollar capitalization of the
Maxim, Blériot, Grahame-White firm.
According to the New York Sun, over $500,000 has been subscribed
for aviation prizes in 1911. The most valuable prizes are for new
records in cross-country flights. The Paris Journal has offered
$70,000 for the best speed in a circling race from Paris to Berlin,
Brussels, London, and back to Paris—1500 miles. Supplementary
prizes from other sources have increased the total stake in this race
to $100,000. A purse of $50,000 is offered by the London Daily Mail
for the “Circuit of Britain” race, from London up the east coast to
Edinburgh, across to Glasgow, and home by way of the west coast,
Exeter, and the Isle of Wight; a thousand miles, to be completed in
two weeks, beginning July 22, with descents only at predetermined
points. This contest will be open (at an entrance fee of $500) to any
licensee of the International Federation. A German circuit, from
Berlin to Bremen, Magdeburg, Düsseldorf, Aix-la-Chapelle, Dresden,
and back to the starting point, is proposed by the Zeitung am Mittag
of Berlin, a prize of $25,000 having been offered. In this country, a
comparatively small prize has been established for a run from San
Francisco to New York, via Chicago. Besides a meet at Bridgeport,
May 18-20, together with those to be held by several of the colleges
and the ones at Bennings and Chicago, there will be, it is still hoped,
a national tournament at Belmont Park at the end of the same
month. Here probably a dozen aviators will contest in qualification
for the international meet in England, to which three American
representatives should be sent as competitors for the championship
trophy now held by Mr. Grahame-White. It is anticipated that the
chances in the international races favor the French aviators, some of
whom—in particular, Leblanc—have been making sensational records
at Pau. Flights between aviation fields in different cities are the
leading feature in the American program for the year. A trip is
proposed from Washington to Belmont Park, via Atlantic City, the
New Jersey coast, and lower New York bay. The distance is 250
miles and the time will probably be less than that of the best
passenger trains between Washington and New York. If held, this
race will probably take place late in May. It is wisely concluded that
the advancement of aviation depends upon cross-country runs under
good control and at reasonable speeds and heights rather than upon
exhibition flights in enclosures. It is to be hoped that commercial
interests will not be sufficiently powerful to hinder this development.
We shall of course have the usual international championship balloon
race, preceded by elimination contests. From present indications
Omaha is likely to be chosen as the point of departure.
The need for scientific study of aerial problems is recognized. The
sum of $350,000 has been offered the University of Paris to found
an aeronautic institute. In Germany, the university at Göttingen has
for years maintained an aerodynamic laboratory. Lord Rayleigh, in
England, is at the head of a committee of ten eminent scientists and
engineers which has, under the authority of Parliament, prepared a
program of necessary theoretical and experimental investigations in
aerostatics and aerodynamics. Our American colleges have organized
student aviation societies and in some of them systematic instruction
is given in the principles underlying the art. A permanent aeronautic
laboratory, to be located at Washington, D.C., is being promoted.
Aviation as a sport is under the control of the International
Aeronautic Federation, having its headquarters at Paris. Bodies like
the Royal Aero Club of England and the Aero Club of America are
subsidiaries to the Federation. In addition, we have in this country
other clubs, like the Aeronautic Society, the United States
Aeronautical Reserve, etc. The National Council of the Aero Clubs of
America is a sort of supreme court for all of these, having control of
meets and contests; but it has no affiliation with the International
body, which is represented here by the Aero Club of America. The
Canadian Auto and Aero Club supervises aviation in the Dominion.
Aviation has developed new legal problems: problems of liability for
accidents to others; the matter of supervision of airship operators.
Bills to license and regulate air craft have been introduced in at least
two state legislatures.
Schools for instruction in flying as an art or sport are being
promoted. It is understood that the Wright firm is prepared to
organize classes of about a dozen men, supplying an aeroplane for
their instruction. Each man pays a small fee, which is remitted
should he afterward purchase a machine. Mr. Grahame-White, at
Pau, in the south of France, conducts a school of aviation, and the
arrangements are now being duplicated in England. Instruction is
given on Blériot monoplanes and Farman biplanes, at a cost of a
hundred guineas for either. The pupil is coached until he can make a
three-mile flight; meanwhile, he is held partially responsible for
damage and is required to take out a “third-party” insurance policy.
There is no lack of aeronautic literature. Major Squier’s paper in the
Transactions of the American Society of Mechanical Engineers, 1908,
gave an eighteen-page list of books and magazine articles of fair
completeness up to its date; Professor Chatley’s book, Aeroplanes,
1911, discusses some recent publications; the Brooklyn Public
Library in New York issued in 1910 (misdated 1909) a manual of
fourteen pages critically referring to the then available literature, and
itself containing a list of some dozen bibliographies.
 AERIAL WARFARE

The German Emperor Watching the Progress of Aviation

The use of air craft as military auxiliaries is not new. As early as
1812 the Russians, before retreating from Moscow, attempted to
drop bombs from balloons: an attempt carried to success by Austrian
engineers in 1849. Both contestants in our own War of Secession
employed captive and drifting balloons. President Lincoln organized a
regular aeronautic auxiliary staff in which one Lowe held the official
rank of chief aeronaut. This same gentleman (who had accomplished
a reconnaissance of 350 miles in eight hours in a 25,000 cubic foot
drifting balloon) was subjected to adverse criticism on account of a
weakness for making ascents while wearing the formal “Prince
Albert” coat and silk hat! A portable gas-generating plant was
employed by the Union army. We are told that General Stoneman, in
1862, directed artillery fire from a balloon, which was repeatedly
fired at by the enemy, but not once hit. The Confederates were less
amply equipped. Their balloon was a patchwork of silk skirts
contributed (one doubts not, with patriotic alacrity) by the daughters
of the Confederacy.
It is not forgotten that communication between besieged Paris and
the external world was kept up for some months during 1870-71 by
balloons exclusively. Mail was carried on a truly commercial scale:
pet animals and—the anticlimax is unintended—164 persons,
including M. Gambetta, escaped in some sixty-five flights. Balloons
were frequently employed in the Franco-Prussian contest; and they
were seldom put hors de combat by the enemy.
During our war with Spain, aerial craft were employed in at least one
instance, namely, at San Juan, Porto Rico, for reconnoitering
entrenchments. Frequent ascents were made from Ladysmith,
during the Boer war. The balloons were often fired at, but never
badly damaged. Cronje’s army was on one occasion located by the
aid of a British scout-balloon. Artillery fire was frequently directed
from aerial observations. Both sides employed balloons in the epic
conflict between Russia and Japan.
A declaration introduced at the second international peace
conference at the Hague proposed to prohibit, for a limited period,
the discharge of projectiles or explosives from flying machines of any
sort. The United States was the only first-class power which
endorsed the declaration. It does not appear likely, therefore, that
international law will discountenance the employment of aerial craft
in international disputes. The building of airships goes on with
increasing eagerness. Last year the Italian chamber appropriated
$5,000,000 for the construction and maintenance of flying machines.
A press report dated February 4 stated that a German aeronaut had
been spending some weeks at Panama, studying the air currents of
the Canal Zone. No flying machine may in Germany approach more
closely than within six miles of a fort, unless specially licensed. At
the Krupp works in Essen there are being tested two new guns for
shooting at aeroplanes and dirigibles. One is mounted on an
armored motor truck. The other is a swivel-mounted gun on a flat-
topped four-wheeled carriage.
The United States battleship Connecticut cost $9,000,000. It
displaces 18,000 tons, uses 17,000 horse-power and 1000 men, and
makes twenty miles an hour. An aeroplane of unusual size with
nearly three times this speed, employing from one to three men with
an engine of 100 horse-power, would weigh one ton and might cost
$5000. A Dreadnought costs $16,000,000, complete, and may last—
it is difficult to say, but few claim more than ten years. It
depreciates, perhaps, at the rate of $2,000,000 a year. Aeroplanes
built to standard designs in large quantities would cost certainly not
over $1000 each. The ratio of cost is 16,000 to 1. Would the largest
Dreadnought, exposed unaided to the attack of 16,000 flying
machines, be in an entirely enviable situation?
An aeroplane is a fragile and costly thing to hazard at one blow: but
not more fragile or costly than a Whitehead torpedo. The aeroplane
soldier takes tremendous risks; but perhaps not greater risks than
those taken by the crew of a submarine. There is never any lack of
daring men when daring is the thing needed.
All experience goes to show that an object in the air is hard to hit.
The flying machine is safer from attack where it works than it is on
the ground. The aim necessary to impart a crippling blow to an
aeroplane must be one of unprecedented accuracy. The dirigible
balloon gives a larger mark, but could not be immediately crippled
by almost any projectile. It could take a good pounding and still get
away. Interesting speculations might be made as to the outcome of
an aerial battle between the two types of craft. The aeroplane might
have a sharp cutting beak with which to ram its more cumbersome
adversary, but this would involve some risk to its own stability: and
the balloon could easily escape by a quick ascent. It has been
suggested that each dirigible would need an aeroplane escort force
for its defense against ramming. Any collision between two opposing
heavier-than-air machines could not, it would seem, be other than
disastrous: but perhaps the dirigible could rescue the wrecks.
Possibly gas-inflated life buoys might be attached to the individual
combatants. In the French manœuvers, a small aeroplane circled the
dirigible with ease, flying not only around it, but in vertical circles
over and under it.

7.5 Centimeter German Automatic Gun for Attacking Airships

(From Brewer’s Art of Aviation)
The French war office has exploited both types of machine. In
Germany, the dirigible has until recently received nearly all the
attention of strategists: but the results of a recent aerial war game
have apparently suggested a change in policy, and the Germans are
now, without neglecting the balloon, actively developing its heavier-
than-air competitor. England seems to be muddled as to its aerial
policy, while the United States has been waiting and for the most
part doing nothing. Now, however, the mobilizations in Texas have
been associated with a considerable amount of aeroplane
enthusiasm. A half-dozen machines, it is expected, will soon be
housed in the aerodrome at San Antonio. Experiments are
anticipated in the carrying of light ammunition and emergency
supplies, and one of the promised manœuvers is to be the locating
of concealed bodies of troops by air scouts. Thirty army officers are
to be detailed for aeroplane service this year; five training schools
are to be established.
If flying machines are relatively unsusceptible to attack, there is also
some question as to their effectiveness in attack. Rifles have been
discharged from moving balloons with some degree of accuracy in
aim; but long-range marksmanship with any but hand weapons
involves the mastery of several difficult factors additional to those
present in gunnery at sea. The recoil of guns might endanger
stability; and it is difficult to estimate the possible effects of a
powerful concussion, with its resulting surges of air, in the
immediate vicinity of a delicately balanced aerial vessel.
But aside from purely combative functions, air craft may be
superlatively useful as messengers. To send despatches rapidly and
without interference, or to carry a general 100 miles in as many
minutes—these accomplishments would render impossible the
romance of a “Sheridan’s Ride,” but might have a romance of their
own. With the new sense added to human equipment by wireless
communication, the results of observations may be signaled to
friends over miles of distance without intervening permanent
connections of however fragile a nature.
Flying machines would seem to be the safest of scouts. They could
pass over the enemy’s country with as little direct danger—perhaps
as unobserved—as a spy in disguise; yet their occupants would
scarcely be subjected to the penalty accompanying discovery of a
spy. They could easily study the movements of an opposing armed
force: a study now frequently associated with great loss of life and
hampering of effective handling of troops. They could watch for
hostile fleets with relatively high effectiveness (under usual
conditions), commanding distant approaches to a long coast line at
slight cost. From their elevated position, they could most readily
detect hostile submarines threatening their own naval fleet.
Maximum effective reconnaissance in minimum time would be their
chief characteristic: in fact, the high speeds might actually constitute
an objection, if they interfered with thorough observation. But if air
craft had been available at Santiago in 1898, Lieutenant Blue’s
expedition would have been unnecessary, and there would have
been for no moment any doubt that Admiral Cervera’s fleet was
actually bottled up behind the Morro. No besieged fortress need any
longer be deprived of communication with—or even some medical or
other supplies from—its friends. Suppose that Napoleon had been
provided with a flying machine at Elba—or even at St. Helena!
The applications to rapid surveying of unknown ground that have
been suggested as possible in civil life would be equally possible in
time of war. Even if the scene of conflict were in an unmapped
portion of the enemy’s territory, the map could be quickly made, the
location of temporary defenses and entrenchments ascertained, and
the advantage of superior knowledge of the ground completely
overcome prior to an engagement. The searchlight and the compass
for true navigation on long flights over unknown country would be
the indispensable aids in such applications.
During the current mobilization of the United States Army at Texas, a
dispatch was carried 21 miles on a map-and-compass flight, the
round trip occupying less than two hours and being made without
incident. The machine flew at a height of 1500 feet and was sighted
several miles off.
A dirigible balloon, it has been suggested, is comparatively safe
while moving in the air, but is subjected to severe strains when
anchored to the ground, if exposed. It must have either safe harbors
of refuge or actual shelter buildings—dry docks, so to speak. In an
enemy’s country a ravine or even a deep railway cut might answer in
an emergency, but the greatest reliance would have to be placed on
quick return trips from a suitable base. The balloon would be,
perhaps, a more effective weapon in defense than in attack. Major
Squier regards a flying height of one mile as giving reasonable
security against hostile projectiles in the daytime. A lower elevation
would be sufficient at night. Given a suitable telephotographic
apparatus, all necessary observations could easily be made from this
altitude. Even in the enemy’s territory, descent to the earth might be
possible at night under reasonably favorable conditions. Two sizes of
balloon would seem to be indicated: the scouting work described
would be done by a small machine having the greatest possible
radius of action. Frontiers would be no barrier to it. Sent from
England in the night it could hover over a Kiel canal or an island of
Heligoland at sunrise, there to observe in most leisurely fashion an
enemy’s mobilizations.
 German Gun for Shooting at Aeroplanes
(From Brewer’s Art of Aviation)
At the London meeting of the Institute of Naval Architects, in April,
1911, the opinion was expressed that the only effective way of
meeting attack from a flying machine at sea would be by a counter-
attack from the same type of craft. The ship designers concluded
that the aeroplane would no more limit the sizes of battleships than
the torpedo has limited them.
For the more serious work of fighting, larger balloons would be
needed, with net carrying capacities perhaps upward from one ton.
Such a machine could launch explosives and combustibles against
the enemy’s forts, dry docks, arsenals, magazines, and battleships.
It could easily and completely destroy his railroads and bridges;
perhaps even his capital itself, including the buildings housing his
chief executive and war office staff. Nothing—it would seem—could
effectually combat it save air craft of its own kind. The battles of the
future may be battles of the air.
There are of course difficulties in the way of dropping missiles of any
great size from flying machines. Curtiss and others have shown that
accuracy of aim is possible. Eight-pound shrapnel shells have been
dropped from an aeroplane with measurably good effect, without
upsetting the vessel; but at best the sudden liberation of a
considerable weight will introduce stabilizing and controlling
difficulties. The passengers who made junketing trips about Paris on
the Clément-Bayard complained that they were not allowed to throw
even a chicken-bone overboard! But it does not seem too much to
expect that these purely mechanical difficulties will be overcome by
purely mechanical remedies. An automatic venting of a gas ballonet
of just sufficient size to compensate for the weight of the dropped
shell would answer in a balloon: a similar automatic change in
propeller speed and angle of planes would suffice with the
aeroplane. There is no doubt but that air craft may be made efficient
agents of destruction on a colossal scale.
 Santos-Dumont Circling the Eiffel Tower
(From Walker’s Aerial Navigation)
A Swedish engineer officer has invented an aerial torpedo,
automatically propelled and balanced like an ordinary submarine
torpedo. It is stated to have an effective radius of three miles while
carrying two and one-half pounds of explosive at the speed of a
bullet. One can see no reason why such torpedoes of the largest size
are not entirely practicable: though much lower speeds than that
stated should be sufficient.
According to press reports, the Krupps have developed a non-
recoiling torpedo, having a range exceeding 5000 yards. The
percussion device is locked at the start, to prevent premature
explosion: unlocking occurs only after a certain velocity has been
attained.
Major Squier apparently contends that the prohibition of offensive
aerial operations is unfair, unless with it there goes the reciprocal
provision that a war balloon shall not be fired at from below. Again,
there seems to be no good reason why aerial mines dropped from
above should be forbidden, while submarine mines—the most
dangerous naval weapons—are allowed. Modern strategy aims to
capture rather than to destroy: the manœuvering of the enemy into
untenable situations by the rapid mobilization of troops being the
end of present-day highly organized staffs. Whether the dirigible
(certainly not the aeroplane) will ever become an effective vehicle
for transport of large bodies of troops cannot yet be foreseen.
Differences in national temper and tradition, and the conflict of
commercial enterprise, perhaps the very recentness of the growth of
a spirit of national unity on the one hand, are rapidly bringing the
two foremost powers of Europe into keen competition: a competition
which is resulting in a bloodless revolution in England, necessitated
by the financial requirements of its naval program. Germany, by its
strategic geographical position, its dominating military organization,
and the enforced frugality, resourcefulness, and efficiency of its
people, possesses what must be regarded as the most invincible
army in the world. Its avowed purpose is an equally invincible navy.
Whether the Gibraltar-Power can keep its ascendancy may well be
doubted. The one doubtful—and at the same time perhaps hopeful—
factor lies in the possibilities of aerial navigation.
 Latham, Farman, and Paulhan
If one battleship, in terms of dollars, represents 16,000 airships, and
if one or a dozen of the latter can destroy the former—a feat not
perhaps beyond the bounds of possibility—if the fortress that
represents the skill and labor of generations may be razed by
twoscore men operating from aloft, then the nations may beat their
spears into pruning-hooks and their swords into plowshares: then
the battle ceases to hinge on the power of the purse. Let war be
made so costly that nations can no more afford it than sane men can
wrestle on the brink of a precipice. Let armed international strife be
viewed as it really is—senseless as the now dying duello. Let the
navy that represents the wealth, the best engineering, the highest
courage and skill, of our age, be powerless at the attack of a swarm
of trifling gnats like Gulliver bound by Lilliputians—what happens
then? It is a reductio ad absurdum. Destructive war becomes so
superlatively destructive as to destroy itself.
There is only one other way. Let the two rival Powers on whom the
peace of the world depends settle their difficulties—surely the earth
must be big enough for both!—and then as one would gently but
firmly take away from a small boy his too destructive toy rifle, spike
the guns and scuttle the ships, their own and all the rest, leaving to
some unambitious and neutral power the prosaic task of policing the
world. Here is a work for red blood and national self-consciousness.
If war were ever needed for man’s best development, other things
will answer now. The torn bodies and desolated homes of millions of
men have paid the price demanded. No imaged hell can surpass the
unnamed horrors that our fathers braved.
“Enforced disarmament!” Why not? Force (and public opinion) have
abolished private duels. Why not national duels as well? Civilization’s
control of savagery always begins with compulsion. For a generation,
no first-class power has had home experience in a serious armed
conflict. We should not willingly contemplate such experience now.
We have too much to do in the world to fight.

The writer has felt some hesitancy in letting these words stand as
the conclusion of a book on flying machines: but as with the old
Roman who terminated every oration with a defiance of Carthage,
the conviction prevails that no other question of the day is of
comparable importance; and on a matter of overwhelming
consequence like this no word can ever be out of place. The five
chief powers spent for war purposes (officially, as Professor Johnson
puts it, for the “preservation of peace”) about $1,000,000,000 in the
year 1908. In the worst period of the Napoleonic operations the
French military and naval budget was less than $100,000,000
annually. Great Britain, on the present peace footing, is spending for
armament more rapidly than from 1793 to 1815. The gigantic “War
Welcome to our website – the ideal destination for book lovers and
knowledge seekers. With a mission to inspire endlessly, we offer a
vast collection of books, ranging from classic literary works to
specialized publications, self-development books, and children's
literature. Each book is a new journey of discovery, expanding
knowledge and enriching the soul of the reade

Our website is not just a platform for buying books, but a bridge
connecting readers to the timeless values of culture and wisdom. With
an elegant, user-friendly interface and an intelligent search system,
we are committed to providing a quick and convenient shopping
experience. Additionally, our special promotions and home delivery
services ensure that you save time and fully enjoy the joy of reading.

Let us accompany you on the journey of exploring knowledge and

personal growth!

textbookfull.com