We investigate methods for manipulating classifier explanations while keeping the predictions unchanged. Our focus is on using a sparse attack, which seeks to alter only a minimal number of input features. We present an efficient and novel algorithm for computing sparse perturbations that alter the explanations but keep the predictions unaffected. We demonstrate that our algorithm, compared to PGD attacks with $\ell_{0}$ constraint, generates sparser perturbations while resulting in greater discrepancies between original and manipulated explanations. Moreover, we demonstrate that it is also possible to conceal the attribution of the k most significant features in the original explanation by perturbing fewer than k features of the input data. We present results for both image and tabular datasets, and emphasize the significance of sparse perturbation-based attacks for trustworthy model building in high-stakes applications. Our research reveals important vulnerabilities in explanation methods that should be taken into account when developing reliable explanation methods. Code can be found at https://github.com/ahmadajal/sparse_expl_attacks

Explanation methods highlight the importance of the input features in taking a predictive decision, and represent a solution to increase the transparency and trustworthiness in machine learning and deep neural networks (DNNs). However, explanation methods can be easily manipulated generating misleading explanations particularly under visually imperceptible adversarial perturbations. Recent work has identified the decision surface geometry of DNNs as the main cause of this phenomenon. To make explanation methods more robust against adversarially crafted perturbations, recent research has promoted several smoothing approaches. These approaches smooth either the explanation map or the decision surface.

Additive models are regression methods which model the response variable as the sum of univariate transfer functions of the input variables. Key benefits of additive models are their accuracy and interpretability on many real-world tasks. Additive models are however not adapted to problems involving a large number (e.g., hundreds) of input variables, as they are prone to overfitting in addition to losing interpretability. In this paper, we introduce a novel framework for applying additive models to a large number of input variables. The key idea is to reduce the task dimensionality by deriving a small number of new covariates obtained by linear combinations of the inputs, where the linear weights are estimated with regard to the regression problem at hand. The weights are moreover constrained to prevent overfitting and facilitate the interpretation of the derived covariates. We establish identifiability of the proposed model under mild assumptions and present an efficient approximate learning algorithm. Experiments on synthetic and real-world data demonstrate that our approach compares favorably to baseline methods in terms of accuracy, while resulting in models of lower complexity and yielding practical insights into high-dimensional real-world regression tasks. Our framework broadens the applicability of additive models to high-dimensional problems while maintaining their interpretability and potential to provide practical insights.

This paper focuses on side information (SI) refinement in Wyner-Ziv video coding and proposes to exploit the intrinsic property of channel coding for improving the joint decoding performance. In this paper, we propose to use syndrome and information bits from the encoder to help the decoder in refining the SI. We use extrinsic information transfer (EXIT) chart analysis to deduce the mutual information variation in LDPC iterative decoding during the SI refinement process. The objective is to obtain the same decoding quality under lower coding rates. Simulation results demonstrate the effectiveness of the proposed solution.

Technologies that use the Internet network to deliver voice communications have the potential to reduce costs and improve access to communications services around the world. However, these new technologies pose several challenges in terms of confidentiality of the conversations and anonymity of the conversing parties. Call authentication and encryption techniques provide a way to protect confidentiality, while anonymity is typically preserved by an anonymizing service (anonymous call). This work studies the feasibility of revealing pairs of anonymous and encrypted conversing parties (caller/callee pair of streams) by exploiting the vulnerabilities inherent to VoIP systems. In particular, by exploiting the aperiodic inter-departure time of VoIP packets, we can trivialize each VoIP stream into a binary time-series. We first define a simple yet intuitive metric to gauge the correlation between two VoIP binary streams. Then we propose an effective technique that progressively pairs conversing parties with high accuracy and in a limited amount of time. Our metric and method are justified analytically and validated by experiments on a very large standard corpus of conversational speech. We obtain impressively high pairing accuracy that reaches 97% after 5 minutes of voice conversations.

Summary form only given. We address the design of any number of balanced descriptions with multiple description scalar quantizers (MDSQ), using fast index assignment methods. Such systems proceed in two steps, scalar quantization and index assignment, that map the quantized value to an N-tuple of quantization indices, to be sent over N channels. We address the specific balanced scenario, where all descriptions have equal rates and where any subset of k out of N descriptions induces the same distortion. We propose two simple index assignment schemes for uniform sources, that are able to generate any number, N (greater than 2), of such balanced descriptions, at any coding rate. The case of Gaussian distributions is also addressed using companding.

Summary form only given. The paper presents a low rate progressive 3D mesh compression scheme for simple, genus-zero 3D objects. The proposed scheme is based on signal representation using redundant expansions on the 2D-sphere. First, generic input data is re-sampled as a function on the 2D-sphere, and the signal value for each point on the regular grid is obtained by performing nearest neighbor interpolation within four points from the initial 3D model. The model representation is then constructed using a matching pursuit algorithm, with an over-complete dictionary of atoms, defined on a sphere. In order to capture the particular characteristics of the 3D models efficiently, we propose a dictionary construction based on two generating functions, a Gaussian to capture low-frequency components, and a modified combination of a Gaussian and its second derivative to capture high-frequency components of the input signal. Compared to state-of-the-art encoders, our method has been shown to offer very good compression efficiency, but the performance is limited by the resampling step that maps the input model on the 2D-sphere. Matching pursuit has, however, the advantage of providing an intrinsically progressive scheme, that is also very flexible.