New adversarial decoy technique bypasses Vision Transformer defenses

Researchers have developed a new method called adversarial decoys to bypass attention-based defenses in Vision Transformers (ViTs). These decoys are independently optimized image patches that redirect the attention mechanism, and consequently the defenses, away from the actual adversarial regions. This technique decouples the misclassification objective from the defense evasion, making it attack-agnostic and easily integrable with existing adversarial patch attacks. Experiments on ImageNet demonstrate that decoys can effectively misdirect attention scores while maintaining significant attack effectiveness, highlighting a limitation in using attention magnitude to detect adversarial relevance. AI

IMPACT This research highlights a vulnerability in current defenses for Vision Transformers, potentially requiring new methods to ensure model robustness against adversarial attacks.

RANK_REASON The cluster contains a research paper detailing a new method for attacking AI models.

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• CORE Recommender
• DagsHub
• Giulia Marchiori Pietrosanti
• Gotit.pub
• Hugging Face
• ImageNet
• Influence Flower
• ScienceCast
• Vision Transformers
• Vít

• paper
• safety

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →