Adversarial Robustness is at Odds with Lazy Training

Autor:	Wang, Yunjuan, Ullah, Enayat, Mianjy, Poorya, Arora, Raman
Rok vydání:	2022
Předmět:	Computer Science - Machine Learning Computer Science - Cryptography and Security Statistics - Machine Learning
Druh dokumentu:	Working Paper
Popis:	Recent works show that adversarial examples exist for random neural networks [Daniely and Schacham, 2020] and that these examples can be found using a single step of gradient ascent [Bubeck et al., 2021]. In this work, we extend this line of work to "lazy training" of neural networks -- a dominant model in deep learning theory in which neural networks are provably efficiently learnable. We show that over-parametrized neural networks that are guaranteed to generalize well and enjoy strong computational guarantees remain vulnerable to attacks generated using a single step of gradient ascent. Comment: NeurIPS 2022
Databáze:	arXiv
Externí odkaz:	http://arxiv.org/abs/2207.00411 Zobrazit plný text záznamu View this record from Arxiv