Adversarial Robustness is at Odds with Lazy Training

Abstract

Recent works show that random neural networks are vulnerable against adversarial attacks [Daniely and Schacham, 2020] and that such attacks can be easily found using a single step of gradient descent [Bubeck et al., 2021]. In this work, we take it one step further and show that a single gradient step can find adversarial examples for networks trained in the so-called lazy regime. This regime is interesting because even though the neural network weights remain close to the initialization, there exist networks with small generalization error, which can be found efficiently using first-order methods. Our work challenges the model of the lazy regime, the dominant regime in which neural networks are provably efficiently learnable. We show that the networks trained in this regime, even though they enjoy good theoretical computational guarantees, remain vulnerable to adversarial examples. To the best of our knowledge, this is the first work to prove that such well-generalizable neural networks are still vulnerable to adversarial attacks.