2019/03/24 by Vinay Kyatham, Mayank Mishra, Kyatham, Vinay +7
Computer Science · #Adversarial Robustness in Machine Learning #Anomaly Detection Techniques and Applications #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Generative Adversarial Networks and Image Synthesis #Machine Learning (cs.LG)
paper · pdf · doi:10.48550/arxiv.1903.09940
openalex publication_date 2019/03/24 · openalex created_date 2022/07/29 · openalex updated_date 2026/07/28
Despite their tremendous success in modelling high-dimensional data\nmanifolds, deep neural networks suffer from the threat of adversarial attacks -\nExistence of perceptually valid input-like samples obtained through careful\nperturbation that lead to degradation in the performance of the underlying\nmodel. Major concerns with existing defense mechanisms include\nnon-generalizability across different attacks, models and large inference time.\nIn this paper, we propose a generalized defense mechanism capitalizing on the\nexpressive power of regularized latent space based generative models. We design\nan adversarial filter, devoid of access to classifier and adversaries, which\nmakes it usable in tandem with any classifier. The basic idea is to learn a\nLipschitz constrained mapping from the data manifold, incorporating adversarial\nperturbations, to a quantized latent space and re-map it to the true data\nmanifold. Specifically, we simultaneously auto-encode the data manifold and its\nperturbations implicitly through the perturbations of the regularized and\nquantized generative latent space, realized using variational inference. We\ndemonstrate the efficacy of the proposed formulation in providing resilience\nagainst multiple attack types (black and white box) and methods, while being\nalmost real-time. Our experiments show that the proposed method surpasses the\nstate-of-the-art techniques in several cases.\n