**Abstract:**

Maliciously manipulated inputs for attacking machine learning methods–in particular deep neural networks–are emerging as a relevant issue for the security of recent artificial intelligence technologies, especially in computer vision. In this paper, we focus on attacks targeting image classifiers implemented with deep neural networks, and we propose a method for detecting adversarial images which focuses on the trajectory of internal representations (ie hidden layers neurons activation, also known as deep features) from the very first, up to the last. We argue that the representations of adversarial inputs follow a different evolution with respect to genuine inputs, and we define a distance-based embedding of features to efficiently encode this information. We train an LSTM network that analyzes the sequence of deep features embedded in a distance space to detect adversarial examples. The results of our preliminary experiments are encouraging: our detection scheme is able to detect adversarial inputs targeted to the ResNet-50 classifier pretrained on the ILSVRC’12 dataset and generated by a variety of crafting algorithms.

**Keywords:** adversarial examples; distance spaces; deep features; machine learning security

**Abstract:**

Relational reasoning in Computer Vision has recently shown impressive results on visual question answering tasks. On the challenging dataset called CLEVR, the recently proposed Relation Network (RN), a simple plug-and-play module and one of the state-of-the-art approaches, has obtained a very good accuracy (95.5%) answering relational questions. In this paper, we define a sub-field of Content-Based Image Retrieval (CBIR) called Relational-CBIR (R-CBIR), in which we are interested in retrieving images with given relationships among objects. To this aim, we employ the RN architecture in order to extract relationaware features from CLEVR images. To prove the effectiveness of these features, we extended both CLEVR and Sort-of-CLEVR datasets generating a ground-truth for R-CBIR by exploiting relational data embedded into scene-graphs. Furthermore, we propose a modification of the RN module–a two-stage Relation Network (2S-RN)–that enabled us to extract relation-aware features by using a preprocessing stage able to focus on the image content, leaving the question apart. Experiments show that our RN features, especially the 2S-RN ones, outperform the RMAC state-of-the-art features on this new challenging task.

**Keywords:** CLEVR, Content-Based Image Retrieval, Deep Learning, Relational Reasoning, Relation Networks, Deep Features

The continuous innovation of satellite payloads is leading to an increasing demand of data-rate for on-board satellite networks. In particular, modern optical detectors generate and need to transfer data at more than 1 Gbps, a speed that cannot be satisfied with standardized technologies such as SpaceWire. To fill this gap, the European Space Agency (ESA) is supporting the development of a new high-speed link standard, SpaceFibre. SpaceFibre provides a data-rate higher than 6.25 Gbps, together with the possibility to use multiple Virtual Channels running over the same physical link, each one configurable with flexible Quality of Service parameters. These features make a SpaceFibre network very appealing but also complex to set up in order to achieve the desired end-to-end requirements. To help this process, a Simulator for HIgh-speed Network (SHINe) based on the open-source toolkit OMNeT++ has been developed and is presented in this paper. It supports the simulation of SpaceFibre and SpaceWire protocols in order to help both the final steps of the standardization process and the system engineers in the setup and test of new networks. SHINe allows to precisely simulate common network metrics, such as latency and bandwidth usage, and it can be connected to real hardware in a Hardware-in-the-Loop configuration.

]]>The vast majority of studies on IP geolocation focuses on localizing the end-users, and little attention has been devoted to localizing the elements of the Internet infrastructure, i.e. the routers and servers that make the Internet work. In this paper we study the maximum theoretical accuracy that can be achieved by a geolocation approach aimed at geolocating the Internet infrastructure. In particular, we study the effects on localization accuracy produced by the position of landmarks and by the strategy followed for their enrollment. We compare two main approaches: the first is more centralized and controlled, and uses well-connected machines belonging to the infrastructure as landmarks; the second is more distributed and scalable, and is based on landmarks positioned at the edge of the network. The study is based on an extensive set of measurements collected using the RIPE Atlas platform. Results show that uniform and widespread diffusion of landmarks can be as important as their measurement accuracy. The study is carried out at both worldwide and regional scale, including regions that were scarcely observed in the past. Results highlight that the geographical characteristics of Internet paths are dependent on the considered region, thus suggesting the use of specifically calibrated models. Finally, the study shows that geolocating IP infrastructure with active measurements is feasible in terms of precision and scalability of the overall system.

]]>URSI EM Theory Symposium, EMTS 2019, San Diego, CA, 27 – 31 May 2019

A comparison between two different uniform asymptotic high-frequency procedures for the evaluation of a typical diffraction integral is presented in this paper. In particular, attention is focused on the specific case of evanescent wave diffraction from a straight wedge [1]. The two procedures are the Pauli-Clemmow (PC) method [2], [3] and the Van der Waerden (VW) method [4], [5]. As well known, the usual leading term of the PC method is not able to provide the proper discontinuity compensation when the poles cross the Steepest Descent Path (SDP) away from the saddle point. However, by considering all higher order terms in the PC asymptotic expansion of the diffraction integral [6], it is shown that each higher order term provides a contribution of order K −1/2 . By suitably collecting all these terms of order K −1/2 , a modified leading term of the PC method is obtained, which results in a compact expression consisting of the standard UTD multiplicative form [1], plus a UTD slope-like correction term [7]. It can be easily demonstrated that the above modified leading term of the PC method exactly coincides with the usual leading term of the VW method. Consequently, in all those cases where the usual leading term of the PC method fails to be uniform, the UTD slope-like correction can be added to obtain a uniform solution.

The modified leading term of the PC method shows an apparent advantage from a numerical point of view, consisting in the fact that its first term coincides with the standard UTD multiplicative form. Moreover, this form has shown to work surprising well in the case of diffraction of a Complex Source Beam by a straight wedge [8], in most case without any need for the UTD slope-like correction.

[1] T.B.A Senior and J.L. Volakis, Approximate Boundary Conditions in Electromagnetics, IEE Electromagnetic Waves Series, The Institution of Electrical Engineers, London, United Kingdom, 1995, pp. 332-336.

[2] P. C. Clemmow, “Some Extensions of the Method of Integration by Steepest Descent,” Quart. J. Mech. Appl. Math., vol. 3, Jan. 1950, pp. 241-256.

[3] P. C. Clemmow, The Plane Wave Spectrum Representation of Electromagnetic Fields, Piscataway, NJ, USA: IEEE Press, pp. 56-58.

[4] B. L. Van der Waerden, “On the Method of Saddle Points,” Appl. Sci. Res. B, vol. 2, no. 1, 1952, pp. 33-45.

[5]. L.B. Felsen and N. Marcuvitz, Radiation and Scattering of Waves, Englewood Cliffs, NJ, USA: Prentice-Hall, 1973.

[6] C. Gennarelli and L. Palumbo, “A Uniform Asymptotic Expansion of a Typical Diffraction Integral with Many Coalescing Single Pole Singularities and a First Order Saddle Point,” IEEE Trans. Antennas Propag., vol. AP-32, no. 10, Oct. 1984, pp. 1122-1124.

[7] R. G: Kouyoumjian, G. Manara, P. Nepa, and B. J. E. Taute, “The Diffraction of an Inhomogeneous Plane Wave by a Wedge,” Radio Science, vol. 31, no. 6, Nov./Dec. 1996, pp. 1387-1397.

[8] H.-T. Chou, P. H. Pathak, Y. Kim, and G. Manara, “On Two Alternative Uniformly Asymptotic Procedures for Analyzing the High-Frequency Diffraction of a Complex Source Beam by a Straight Wedge,” IEEE Trans. Antennas Propag., vol. AP-66, no. 7, July 2018, pp. 3631-3641.

]]>

URSI AP-RASC 2019, New Delhi, India, 09 - 15 March 2019

A Uniform Geometrical Theory of Diffraction (UTD) solution for inhomogeneous plane wave diffraction by a Perfect Electrically Conducting (PEC) wedge has been presented in [1]. More recently, this solution has been extended to analyze the case of an incident field radiated by a source located in the complex space (Complex Source Beam, CSB) [2], [3]. This solution has shown to be very accurate when compared with reference data obtained by a rigorous multipole expansion of the field. The purpose of this communication is to further investigate some interesting properties of the extended UTD solution, opening the way to a better understanding of the physical phenomena connected to an evanescent behavior of the waves impinging on the edge. In the context of high-frequency techniques, the UTD solution for CSB diffraction in [2] has been written in a simple and compact form which includes the incident, the reflected and the diffracted ray contributions, so that it can be directly applied to calculate the scattering from more complex geometries with edges [4], [5]. Here, the analysis is oriented to gain a better understanding of the mechanism which controls discontinuity compensations at shadow and reflection boundaries. Also, attention is focused on the identification of shadow and reflection boundaries, as well as in determining both the shape and the extension of the pertinent transition regions. All these investigations are verified through comparisons with a rigorous multipole expansion solution.

1. R.G. Kouyoumjian, G. Manara, P. Nepa, and B.J.E. Taute, “The diffraction of an inhomogeneous plane wave by a wedge,” Radio Science, vol. 31, Nov./Dec. 1996, pp. 1387-1397.

2. S. Terranova, G. Manara, L. Klinkenbusch, A Physical Insight into Complex-Source Beam Diffraction by a Wedge,” 2nd URSI Atlantic Meeting, May 28 – June 1st, 2018, Gran Canaria, Spain.

3. L. B. Felsen, “Complex source-point solutions of the field equations and their relation to the propagation and scattering of Gaussian beams,” in Symposia Matematica, Istituto Nazionale di Alta Matematica, London, U. K., Academic, 1976, Vol. XVIII, pp. 40 – 56.

4. M. Katsav, E. Heyman, and L. Klinkenbusch, “Beam diffraction by a wedge: exact and complex ray solutions,” IEEE Transactions on Antennas and Propagation, Vol. 62, No. 7, pp. 3731 – 3740, July 2014.

5. H.-T. Chou, P. H. Pathak, Y. Kim, and G. Manara, “On Two Alternative Uniformly Asymptotic Procedures for Analyzing the High-Frequency Diffraction of a Complex Source Beam by a Stright Edge,” IEEE Transactions on Antennas and Propagation, Vol. 66, No. 7, pp. 3631-3642, July 2018.

]]>https://www.mdpi.com/2079-9292/8/1/35

Published in MDPI Electronics as part of the Special Issue RFID, WPT and Energy Harvesting

In this paper, a low-cost chipless reader for detecting depolarizing tags is described. The reader operates in the frequency band (2–2.5) GHz, and it is compact and integrated in a single board. The reader architecture and its transmitting and receiving antennas are presented. Reader antennas comprise of two orthogonally placed, E-shaped patches with a decoupling below −35 dB. The reader performance is evaluated on a four-bit tag formed by four obliquely placed dipoles on top of a metallic ground plane.

]]>The process of geolocating Internet hosts from their IP addresses using delay measurements highly depends on the measured speed-of-Internet (SOI). This parameter can differ in different areas of the world due to the circuitousness of Internet paths. In a previous work we have discovered an anomaly in Middle East: the SOI is much slower than in other world regions. In this paper we confirm this anomaly by the means of measurements collected with RIPE Atlas. We analyze the topological and geographical properties of Internet paths in Middle East and compare them to European ones to highlight the reasons of this anomaly. We show that the traffic between two endpoints in Middle East is highly subject to circuitous paths, due to a low incidence of Internet eXchange Points and a scarce interconnection among countries in the region, which inhibits the locality of Internet traffic. In particular, neighbouring countries do not communicate directly, instead Internet paths often traverse even different regions before reaching the destination. This increases delay and circuitousness, and in the end reduces the SOI.

]]>Paper published at ITASEC'19, the Italian Conference on Cybersecurity (http://ceur-ws.org/Vol-2315/paper13.pdf)

]]>This work investigates collaborative aerial transportation by swarms of agents based only on implicit information, enabled by the physical interaction among the agents and the environment. Such a coordinating mechanism in collaborative transportation is a basic skill in groups of social animals. We consider cable-suspended objects transported by a swarm of flying robots and we formulate several hypothesis on the behavior of the overall system which are validated thorough numerical study. In particular, we show that a nonzero internal force reduces to one the number of asymptotically stable equilibria and that the internal force intensity is directly connected to the convergence rate. As such, the internal force represents the cornerstone of a communication-less cooperative manipulation paradigm in swarms of flying robots. We also show how a swarm can achieve a stable transportation despite the imprecise knowledge of the system parameters.

]]>