Reproduce “SAFER: Sharpness Aware layer-selective Finetuning for Enhanced Robustness in vision transformers” 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

Bhavna Gopal, Huanrui Yang, Mark Horton, Yiran Chen

SAFER addresses image-classifier failures caused by designed small changes, which adversarial training can overfit. It ranks transformer layers by how sharply adversarial loss responds to small weight changes, then applies sharpness-aware fine-tuning to roughly 5% of layers while freezing the rest. On DeiT-Tiny/CIFAR-10, it reached 82.36% clean, 68.50% PGD-20, and 50.12% AutoAttack accuracy, but results remain empirical.

Summary by alphaXiv.

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