Federated Learning (FL) enhances data privacy by enabling users to collaboratively train neural networks without sharing raw data. However, FL does not guarantee model privacy because clients share gradients of the model. This paper introduces a novel attack against FL in the regression setting, uncovering a previously unexplored vulnerability. We show that, in regression tasks where labels are continuous real numbers, gradient equations can be reduced to a hidden subset sum problem (HSSP, a cryptographic problem originally studied for integer values). By adapting cryptographic techniques for solving the HSSP, we demonstrate that labels can be accurately recovered, posing a significant privacy risk. Unlike analytical approaches for classification such as iDLG [22] that do not extend to regression, our method is specifically designed for this setting. Through extensive experiments on eight datasets with three different deep learning models, we demonstrate that label recovery achieves very low reconstruction error (e.g., MSE below 0.1 in some scenarios), significantly outperforming prior gradient inversion attacks such as DLG [23].

Multi-Source Domain Adaptation (MSDA) is a challenging scenario where multiple related and heterogeneous source datasets must be adapted to an unlabeled target dataset. Conventional MSDA methods often overlook that data holders may have privacy concerns, hindering direct data sharing. In response, decentralized MSDA has emerged as a promising strategy to achieve adaptation without centralizing clients' data. Our work proposes a novel approach, Decentralized Dataset Dictionary Learning, to address this challenge. Our method leverages Wasserstein barycenters to model the distributional shift across multiple clients, enabling effective adaptation while preserving data privacy. Specifically, our algorithm expresses each client's underlying distribution as a Wasserstein barycenter of public atoms, weighted by private barycentric coordinates. Our approach ensures that the barycentric coordinates remain undisclosed throughout the adaptation process. Extensive experimentation across five visual domain adaptation benchmarks demonstrates the superiority of our strategy over existing decentralized MSDA techniques. Moreover, our method exhibits enhanced robustness to client parallelism while maintaining relative resilience compared to conventional decentralized MSDA methodologies.

In this article, we propose an approach for federated domain adaptation, a setting where distributional shift exists among clients and some have unlabeled data. The proposed framework, FedDaDiL, tackles the resulting challenge through dictionary learning of empirical distributions. In our setting, clients' distributions represent particular domains, and FedDaDiL collectively trains a federated dictionary of empirical distributions. In particular, we build upon the Dataset Dictionary Learning framework by designing collaborative communication protocols and aggregation operations. The chosen protocols keep clients' data private, thus enhancing overall privacy compared to its centralized counterpart. We empirically demonstrate that our approach successfully generates labeled data on the target domain with extensive experiments on (i) Caltech-Office, (ii) TEP, and (iii) CWRU benchmarks. Furthermore, we compare our method to its centralized counterpart and other benchmarks in federated domain adaptation.

Federated Learning (FL) has gained prominence as a decentralized and privacy-preserving paradigm that enables multiple clients to collaboratively train a machine learning model under the supervision of a central server. Instead of centralizing the data, clients keep their data locally and share only model parameters during multiple communication rounds. However, recent attacks, such as gradient reconstruction attacks (GRAs) show privacy issues when an attacker knows the communication of a client. In the literature, these privacy issues are mainly explored when clients compute new parameters using a single gradient descent step on their data (FedSGD) and then send them back to the remote server. In a more realistic scenario, the clients' protocol is based on several gradient descent steps (FedAvg). This protocol adds intermediate computation steps, which are unknown from the attacker, thus making GRAs less successful. In this incremental paper, we conduct exhaustive experiments on four state-of-the-art attacks under the FedAvg protocol, on a very basic and a more complex neural network (ResNet-18) with CIFAR100 dataset. These experiments provide the following results 1) a privacy-utility trade-off analysis, 2) insights on the choice of attacks' hyperparameters, 3) the client's local learning rate has little impact on attacks' effectiveness 4) a proof that the privacy risk is not necessarily decreasing over rounds, contrary to common belief.

Federated Learning is a decentralized framework that enables multiple clients to collaboratively train a machine learning model under the orchestration of a central server without sharing their local data. The centrality of this framework represents a point of failure which is addressed in literature by blockchain-based federated learning approaches. While ensuring a fully-decentralized solution with traceability, such approaches still face several challenges about integrity, confidentiality and scalability to be practically deployed. In this paper we propose Fantastyc, a solution designed to address these challenges that have been never met together in the state of the art.

Robust and accurate path tracking for autonomous vehicle navigation is a complex task, especially when it comes to managing system uncertainties such as inertia, slippage, and action delays. Although model-based controllers are efficient, their performance can be limited by such uncertainties and by the complexity of the gain tuning process. To address this, our study evaluates the effectiveness of four strategies using reinforcement learning (RL) with a controller, to provide either - steering correction, full gain tuning, gain correction, or end-to-end learning without any controller - to improve trajectory tracking. These methods are trained on geometric controllers (Pure Pursuit, Stanley) and model predictive controllers (Romea, EBSF). Our results show that all RL methods improve tracking at high speeds, with steering correction proving the most consistently effective in all cases.

This paper proposes a post-training defense against pattern-triggered backdoor attacks in federated learning contexts. This approach relies first on the server estimating the attack pattern. The server then provides the estimated pattern to the end-users, who use it directly on their local data to mitigate backdoor attacks during inference time. This scheme offers an improvement over the existing approaches by demonstrating robustness to data heterogeneity among users without needing a shared dataset or additional information from users and regardless of the number of malicious clients. Based on extensive comparison with existing state-of-the-art methods on well-known computer vision datasets, the proposed method is shown to succeed in mitigating backdoor attacks while preserving high accuracy on clean inputs.

Federated learning enables different parties to collaboratively build a global model under the orchestration of a server while keeping the training data on clients' devices. However, performance is affected when clients have heterogeneous data. To cope with this problem, we assume that despite data heterogeneity, there are groups of clients who have similar data distributions that can be clustered. In previous approaches, in order to cluster clients the server requires clients to send their parameters simultaneously. However, this can be problematic in a context where there is a significant number of participants that may have limited availability. To prevent such a bottleneck, we propose FLIC (Federated Learning with Incremental Clustering), in which the server exploits the updates sent by clients during federated training instead of asking them to send their parameters simultaneously. Hence no additional communications between the server and the clients are necessary other than what classical federated learning requires. We empirically demonstrate for various non-IID cases that our approach successfully splits clients into groups following the same data distributions. We also identify the limitations of FLIC by studying its capability to partition clients at the early stages of the federated learning process efficiently. We further address attacks on models as a form of data heterogeneity and empirically show that FLIC is a robust defense against poisoning attacks even when the proportion of malicious clients is higher than 50%.

In this paper, we present the first Federated Learning (FL) framework which is secure against both confidentiality and integrity threats from the aggregation server, in the case where the resulting model is not disclosed to the latter. We do so by combining Homomorphic Encryption (HE) and Verifiable Computing (VC) techniques in order to perform a Federated Averaging operator directly in the encrypted domain (by means of HE) and produce formal proofs that the operator was correctly applied (by means of VC). Due to the simplicity of the aggregation function, we are able to ground our approach in additive HE techniques which are highly mature in terms of security and decently efficient. We also introduce a number of optimizations which allows to reach practical execution performances on the larger deep learning models end of the spectrum. The paper also provides extensive experimental results on the FEMNIST dataset demonstrating that the approach preserves the quality of the resulting models at the cost of practically meaningful computing and communication overheads, at least in the cross-silo setting for which higher-end machines can be involved on both the client and server sides.

These last years, artificial neural networks (ANN) have known a renewed interest since efficient training procedures have emerged to learn the so called deep neural networks (DNN), i.e. ANN with at least two hidden layers. In the same time, the computational auditory scene recognition (CASR) problem which consists in estimating the environment around a device from the received audio signal has been investigated. Most of works which deal with the CASR problem have tried to ind well-adapted features for this problem. However, these features are generally combined with a classical classi-ier. In this paper, we introduce DNN in the CASR ield and we show that such networks can provide promising results and perform better than standard classiiers when the same features are used.

This paper depicts the performances of a multi-sensor prototype on explosive vapor detection. The responses of the device in laboratory conditions but also in real life conditions were evaluated. Explosive precursors, explosive compositions and taggants were detected independently and under complex conditions, in presence of interfering compounds. Indeed, an efficient detector should detect small quantities of explosive in a realistic environment. It should also be selective to explosives and exhibit a very low level of false alarms when used in presence of benign substances or objects.

In this paper, a novel gas identification approach based on the Recursive Least Squares (RLS) algorithm is proposed. We detail some adaptations of RLS to be applied to a sensor matrix of several technologies in optimal conditions. The low complexity of the algorithm and its ability to process online samples from multi-sensor make the real-time identification of volatile compounds possible. The effectiveness of this approach to early detect and recognize explosive compounds in the air has been successfully demonstrated on an experimentally obtained dataset.

On one hand, sparse coding, which is widely used in signal processing, consists of representing signals as linear combinations of few elementary patterns selected from a dedicated dictionary. The output is a sparse vector containing few coding coefficients and is called sparse code. On the other hand, Multilayer Perceptron (MLP) is a neural network classification method that learns non linear borders between classes using labeled data examples. The MLP input data are vectors, usually normalized and preprocessed to minimize the inter-class correlation. This article acts as a link between sparse coding and MLP by converting sparse code into convenient vectors for MLP input. This original association assures in this way the classification of any sparse signals. Experimental results obtained by the whole process on trajectories data and comparisons to other methods show that this approach is efficient for signals classification.

Sparse signal decomposition are keys to efficient compression, storage and denoising, but they lack appropriate methods to exploit this sparsity for a classification purpose. Sparse coding methods based on dictionary learning may result in spikegrams, a sparse and temporal representation of signals by a raster of kernel occurrence through time. This paper proposes a method for coupling spike train cost based metrics (from neuroscience) with a spikegram sparse decompositions for clustering multivariate signals. Experiments on character trajectories, recorded by sensors from natural handwriting, prove the validity of the approach, compared with currently available classification performance in literature.

Sparse signal decomposition are keys to efficient compression, storage and denoising, but they lack appropriate methods to exploit this sparcity for a classification purpose. Sparse coding methods based on dictionary learning may result in spikegrams, a sparse and temporal representation of signals by a raster of kernel occurrence through time. This paper proposes a method for coupling spike train cost based metrics (from neuroscience) with a spikegram sparse decompositions for clustering multivariate signals. Experiments on character trajectories, recorded by sensors from natural handwriting, prove the validity of the approach, compared with currently available classification performance in literature.

Classical dictionary learning algorithms (DLA) allow unicomponent signals to be processed. Due to our interest in two-dimensional (2D) motion signals, we wanted to mix the two components to provide rotation invariance. So, multicomponent frameworks are examined here. In contrast to the well-known multichannel framework, a multivariate framework is first introduced as a tool to easily solve our problem and to preserve the data structure. Within this multivariate framework, we then present sparse coding methods: multivariate orthogonal matching pursuit (M-OMP), which provides sparse approximation for multivariate signals, and multivariate DLA (M-DLA), which empirically learns the characteristic patterns (or features) that are associated to a multivariate signals set, and combines shift-invariance and online learning. Once the multivariate dictionary is learned, any signal of this considered set can be approximated sparsely. This multivariate framework is introduced to simply present the 2D rotation invariant (2DRI) case. By studying 2D motions that are acquired in bivariate real signals, we want the decompositions to be independent of the orientation of the movement execution in the 2D space. The methods are thus specified for the 2DRI case to be robust to any rotation: 2DRI-OMP and 2DRI-DLA. Shift and rotation invariant cases induce a compact learned dictionary and provide robust decomposition. As validation, our methods are applied to 2D handwritten data to extract the elementary features of this signals set, and to provide rotation invariant decomposition.

In this paper, we present the main results of the BioSecure Signature Evaluation Campaign (BSEC'2009). The objective of BSEC'2009 was to evaluate different online signature algorithms on two tasks: the first one aims at studying the influence of acquisition conditions (digitizing tablet or PDA) on systems' performance; the second one aims at studying the impact of information content in signatures on systems' performance. In BSEC'2009, the two BioSecure Data Sets DS2 and DS3 are used for tests, both containing data of the same 382 people, acquired respectively on a digitizing tablet and on a PDA. The results of the 12 systems involved in this evaluation campaign are reported and analyzed in detail in this paper. Experimental results reveal a 2.2% EER for skilled forgeries and a 0.51% EER for random forgeries on DS2; and a 4.97% EER for skilled forgeries and a 0.55% EER for random forgeries on DS3.