AdvRefactor: A Resampling-Based Defense Against Adversarial Attacks

摘要

Deep neural networks have achieved great success in many domains. However, they are vulnerable to adversarial attacks, which generate adversarial examples by adding tiny perturbations to legitimate images. Previous studies providing defense mostly focus on modifying DNN models to mitigate adversarial attacks. We propose a resampling-based defense, AdvRefactor, which aims at transforming the inputs of models, and thereby eliminates the adversarial perturbations. We explore two resampling algorithms, proximal interpolation and bilinear interpolation, which cost less, suit more models and are combinable with other defenses. Our evaluation results demonstrate that AdvRefactor can significantly mitigate the adversarial attacks.