Open access
2026
Target-Oriented Adversarial Patches Against Vision–Language Models via Heatmap-Guided Localization
A target-oriented adversarial patch framework that addresses both patch placement and patch optimization, and cross-model evaluation indicates that patches exhibit non-trivial transferability within the CLIP family, particularly when EVA-CLIP is used as the source model.
Xue-Ying Wang, Ding-Yi Lu, Cheng-Ci Hu et al.
· IEEE Access · 0 citations