Attacking only the top 20% high-entropy token positions in vision-language models causes comparable semantic damage and more harmful outputs than global attacks, and these vulnerable tokens transfer across model architectures.
Simlabel: Consistency-guided OOD detection with pretrained vision-language models.CoRR, abs/2501.11485
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
High-Entropy Tokens as Multimodal Failure Points in Vision-Language Models
Attacking only the top 20% high-entropy token positions in vision-language models causes comparable semantic damage and more harmful outputs than global attacks, and these vulnerable tokens transfer across model architectures.