Vision-language model judgments about image realism and generated captions can be shifted by imperceptible perturbations confined to specific spatial frequency bands, even under black-box access.
Towards multimodal disinformation detection by vision-language knowledge in- teraction,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
support 1representative citing papers
citing papers explorer
-
On the Reliability of Vision-Language Models Under Adversarial Frequency-Domain Perturbations
Vision-language model judgments about image realism and generated captions can be shifted by imperceptible perturbations confined to specific spatial frequency bands, even under black-box access.