ai_security▌

← feed

NOTABLE

arXiv: GARD hunts backdoors in CLIP-style encoders with structure-matched adversarial probes

arXiv 2611.14143 observes that backdoored multimodal contrastive encoders show structure-dependent representation sensitivity - global perturbations expose distributed backdoors, local patch perturbations expose localized ones - and builds GARD, a post-hoc defense that needs neither the attack target nor the trigger.

released 2026-11-10 · ingested 2026-11-11T12:15Z

Sources

Primary source

Cite this

APA

AI Security Index. (2026). arXiv: GARD hunts backdoors in CLIP-style encoders with structure-matched adversarial probes. AI Security Index. https://arxiv.org/abs/2611.14143

BibTeX

@misc{aisec-2026-11-11-arxiv-2611-14143-gard-backdoor-defense-clip-encoders,
  title  = {arXiv: GARD hunts backdoors in CLIP-style encoders with structure-matched adversarial probes},
  year   = {2026},
  howpublished = {\url{https://arxiv.org/abs/2611.14143}},
  note   = {AI Security Index item 2026-11-11-arxiv-2611-14143-gard-backdoor-defense-clip-encoders}
}