Reproduce “Probabilistic Margins for Instance Reweighting 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
Qizhou Wang, Feng Liu, Bo Han, Tongliang Liu, Chen Gong, Gang Niu, Mingyuan Zhou, Masashi Sugiyama
MAIL reweights adversarial-training examples using a probabilistic margin: the model’s predicted probability for the true class minus that of its strongest competitor. Smaller margins receive greater weight, prioritizing examples the authors consider more vulnerable. On CIFAR-10 with ResNet-18, MAIL-TRADES reached 50.60 ± 0.22% AutoAttack accuracy versus GAIRAT’s 33.35 ± 0.57%, under the tested white-box 8/255 setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.