Scaling Adversarial Training to Large Perturbation Bounds
This addresses the vulnerability of AI systems to more realistic adversarial attacks beyond typical low-magnitude constraints, though it is incremental as it builds on existing adversarial training methods.
The paper tackles the problem of adversarial robustness for deep neural networks against large perturbation bounds, where real-world adversaries are not limited by low magnitude constraints, and proposes Oracle-Aligned Adversarial Training (OA-AT) to align network predictions with an Oracle, achieving state-of-the-art performance at large bounds (e.g., L-inf bound of 16/255 on CIFAR-10) and outperforming existing defenses at standard bounds (8/255).
The vulnerability of Deep Neural Networks to Adversarial Attacks has fuelled research towards building robust models. While most Adversarial Training algorithms aim at defending attacks constrained within low magnitude Lp norm bounds, real-world adversaries are not limited by such constraints. In this work, we aim to achieve adversarial robustness within larger bounds, against perturbations that may be perceptible, but do not change human (or Oracle) prediction. The presence of images that flip Oracle predictions and those that do not makes this a challenging setting for adversarial robustness. We discuss the ideal goals of an adversarial defense algorithm beyond perceptual limits, and further highlight the shortcomings of naively extending existing training algorithms to higher perturbation bounds. In order to overcome these shortcomings, we propose a novel defense, Oracle-Aligned Adversarial Training (OA-AT), to align the predictions of the network with that of an Oracle during adversarial training. The proposed approach achieves state-of-the-art performance at large epsilon bounds (such as an L-inf bound of 16/255 on CIFAR-10) while outperforming existing defenses (AWP, TRADES, PGD-AT) at standard bounds (8/255) as well.