Reproduce “ExGrad: Beyond Suppression - Explicit Clean-Gradient Subtraction in Adversarial Training” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhiyuan Ye, Xiangyu Zhou, Ji Qi, Hao Zhang, Yi Zhou

ExGrad selects a training update by probing how clean, adversarial-input, and adversarial-weight-perturbation gradients change margins on attacked examples, then recombining them with response-adaptive affine coefficients. On CIFAR-10, allowing a clean-gradient coefficient of −0.1 rather than 0 raised PGD-20 accuracy from 56.44% to 57.12%, while clean accuracy fell. This controlled result does not establish that clean fitting is generally harmful.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.