GARD: Gradient-Guided Adversarial Representation Defense against Backdoors in Multimodal Contrastive Learning Models

Summary

GARD is a post-hoc defense for backdoored multimodal contrastive encoders such as CLIP that needs no knowledge of the attack target or trigger. It optimizes complementary global and local adversarial probes to displace clean-image representations, then realigns them with detached clean anchors while preserving image-text alignment. Across six visual backdoor attacks and two CLIP visual backbones, it reduces attack success rate below 1% in all 12 combinations while keeping competitive clean accuracy.

Mitigation

GARD is the proposed mitigation: a target- and trigger-agnostic post-hoc defense that realigns perturbed representations with detached clean anchors, requiring no additional computation at inference time.

Related items

• HighCVE-2026-63651: LiveDocs is a primary-source-first documentation for AI agents. Prior to version 0.7.0, fetch_docs and introspect…Similar attack · NVD/CVE Database
• HighCVE-2026-125271 - AWS API MCP Server Security Policy Bypass via Incorrect Operation Name NormalizationSimilar attack · AWS Security Bulletins
• InfoLocation Matters: Sparse Latent-Variance Attacks for Face Privacy against Diffusion PersonalizationSimilar attack · Arxiv (cs.RO + cs.CV security)
• MediumCVE-2026-83953 GitHub Copilot for Jetbrains Information Disclosure VulnerabilitySimilar attack · MSRC Security Update Guide
• HighGHSA-ghwv-vrp2-2975: Cocos AI Intra-handshake attested TLS implementation is vulnerable to TOCTOU AttacksSimilar attack · GitHub Advisory Database