MAA boosts the transferability of adversarial image attacks on vision-language models by combining randomly resized sliding crops with multi-layer feature and cross-modal similarity disruption.
Learning to learn transferable attack
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
-
MAA: Meticulous Adversarial Attack against Vision-Language Pre-trained Models
MAA boosts the transferability of adversarial image attacks on vision-language models by combining randomly resized sliding crops with multi-layer feature and cross-modal similarity disruption.