Reproduce “GARD: Gradient-Guided Adversarial Representation Defense against Backdoors in Multimodal Contrastive Learning Models” 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
Yuhong Wang, Yujun Cai, Lei Feng, Songze Li
The authors present GARD, a defense for backdoored CLIP models. It uses global and localized adversarial image probes, then trains representations toward detached clean-image anchors while preserving image-text matching. In ImageNet-1K classification, GARD reports ASR below 1% across six attacks and two visual backbones, with competitive clean accuracy. This result is limited to those evaluations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.