Poisoning attack is one of the attack types commonly studied in the field of adversarial machine learning. The adversary generating poison attacks is assumed to have access to the training process of a machine learning algorithm and aims to prevent the algorithm from functioning properly by injecting manipulative data while the algorithm is being trained. In this work, our focus is on poisoning attacks against linear regression models which target to weaken the prediction power of the attacked regression model. We propose a bilevel optimization problem to model this adversarial process between the attacker generating poisoning attacks and the learner which tries to learn the best predictive regression model. We give an alternative single level optimization problem by benefiting from the optimality conditions of the learner’s problem. A commercial solver is used to solve the resulting single level optimization problem where we generate the whole set of poisoning attack samples at once. Besides, an iterative approach that allows to determine only a portion of poisoning attack samples at every iteration is introduced. The proposed attack strategies are shown to be superior than a benchmark algorithm from the literature by carrying out extensive experiments on two realistic datasets.
Differential privacy is a formal definition of privacy that bounds the maximum acceptable information leakage when a query is performed on sensitive data. To ensure this property, a key technique involves bounding the query's sensitivity (how much input variations affect the output) and adding noise to the result according to this quantity. While prior work like the Fuzz type system focuses on global sensitivity, many useful queries have infinite global sensitivity, restricting the scope of such approaches. This limitation can be addressed by considering a more fine-grained measure: local sensitivity, which quantifies output change for inputs adjacent to a specific dataset. In this article, we introduce Local Fuzz, a type system with dependent coeffects designed to bound the local sensitivity of programs written in a simple functional language. We provide a denotational semantics for this system in the category of extended premetric spaces, leveraging the recently introduced construction of a dependently graded comonad. Finally, we illustrate how Local Fuzz can lead to better differential privacy guarantees than Fuzz, both for mechanisms that rely on global sensitivity and for those that leverage local sensitivity, such as the Propose-Test-Release framework.
In this article, we study the inconsistency of a system of $\max-T$ fuzzy relational equations of the form $A \Box_{T}^{\max} x = b$, where $T$ is a t-norm among $\min$, the product or Lukasiewicz's t-norm. For an inconsistent $\max-T$ system, we directly construct a canonical maximal consistent subsystem (w.r.t the inclusion order). The main tool used to obtain it is the analytical formula which compute the Chebyshev distance $\Delta = \inf_{c \in \mathcal{C}} \Vert b - c \Vert$ associated to the inconsistent $\max-T$ system, where $\mathcal{C}$ is the set of second members of consistent systems defined with the same matrix $A$. Based on the same analytical formula, we give, for an inconsistent $\max-\min$ system, an efficient method to obtain all its consistent subsystems, and we show how to iteratively get all its maximal consistent subsystems.
We propose three floating-point validated algorithms to compute respectively fast inversion, Euclidean division and Hensel lifting over ℂ[[x]][y] . This is the second step (after Bréhard, Poteaux and Soudant in ISSAC 2023) towards a validated numerical Newton–Puiseux algorithm, and will also be useful towards a validated OM-algorithm over ℂ[[x]][y] . Our strategy is simply to first compute a floating-point approximation using the classical algorithm, then to a posteriori validate the result using a Newton-like fixed-point operator. We also provide a prototype Julia implementation of these algorithms and several examples.
This paper presents a novel method for compressing Convolutional Neural Networks (CNNs) to enable efficient deployment on low-capacity devices. The proposed approach combines neural network pruning with reinforcement learning (RL) and graph embedding. Each network is represented as a computational graph, and Graph Convolutional Networks (GCNs) are utilized to learn graph-level embeddings that inform pruning decisions. By applying Proximal Policy Optimization (PPO), we automate the selection of layer-wise pruning ratios, eliminating the need for manual tuning. Experiments on ResNet-34 and VGG-19, trained on the CIFAR-10 dataset, demonstrate that our method achieves up to 80% compression while maintaining or improving model accuracy through post-pruning rewinding. We evaluated both structured and unstructured pruning strategies, analyzing the trade-offs in accuracy, FLOPs, parameter count, and inference time
