README.md

metadata

license: apache-2.0
pipeline_tag: zero-shot-image-classification

Towards Highly Transferable Vision-Language Attack via Semantic-Augmented Dynamic Contrastive Interaction

This repository contains the official implementation of SADCA (Semantic-Augmented Dynamic Contrastive Attack), presented in the paper Towards Highly Transferable Vision-Language Attack via Semantic-Augmented Dynamic Contrastive Interaction.

SADCA is a framework designed to enhance the transferability of adversarial attacks against vision-language pre-training (VLP) models. It progressively disrupts cross-modal alignment through dynamic interactions between adversarial images and texts, using a contrastive learning mechanism involving adversarial, positive, and negative samples to reinforce semantic inconsistency.

Links

• Paper: https://arxiv.org/abs/2603.04839
• GitHub: https://github.com/LiYuanBoJNU/SADCA

Citation

@article{li2026towards,
  title={Towards Highly Transferable Vision-Language Attack via Semantic-Augmented Dynamic Contrastive Interaction},
  author={Li, Yuanbo and Xu, Tianyang and Hu, Cong and Zhou, Tao and Wu, Xiao-Jun and Kittler, Josef},
  journal={arXiv preprint arXiv:2603.04839},
  year={2026}
}