GARD: Gradient-Guided Adversarial Representation Defense Against Backdoors in Multimodal Contrastive Learning Models
Start here
GARD performs post hoc defense against backdoors in multimodal contrastive learning models such as CLIP, without requiring prior knowledge of the attack target or trigger: it optimizes global and local adversarial probes to shift clean image representations, then realigns them with clean anchors. Experiments cover all 12 combinations of 6 backdoor attacks and 2 CLIP vision backbones, reducing attack success rate to below 1% while maintaining competitive clean accuracy, with no additional computation required at inference time.
Generated by AI from reports · Updated 2 days ago
Related research
This event's paper
- GARD: Gradient-Guided Adversarial Representation Defense against Backdoors in Multimodal Contrastive Learning Models
arXiv · 2026-11-10 · Deep reading
Progress timeline
- arXiv: Adversarial Attacks, Backdoors & PoisoningGARD: Gradient-Guided Adversarial Representation Defense Against Backdoors in Multimodal Contrastive Learning Models
GARD is a post-hoc defense method against backdoors in multimodal contrastive learning models such as CLIP, requiring no prior knowledge of the attack target or trigger. The method optimizes complementary global and local adversarial probes to shift clean image representations, then realigns them with clean anchors while maintaining image-text alignment. Across all 12 combinations of 6 visual backdoor attacks and 2 CLIP vision backbones, the attack success rate drops below 1%, clean accuracy remains competitive, and no additional computation is needed at inference time.
Related discussion
Attention trend
New sources per day
- Nov 11: 1 new sources (1 outlets, 0 accounts)
- Nov 12: 0 new sources (0 outlets, 0 accounts)
- Nov 13: 0 new sources (0 outlets, 0 accounts)