New Framework Attacks Vision-Language Models via Dual Modalities

Researchers have introduced Multi-Modal Adversarial Synergy (MMAS), a black-box framework that simultaneously generates adversarial perturbations for both images and text to attack Large Vision-Language Models (LVLMs). The image perturbation uses wavelet-based texture constraints, while the text perturbation is a learnable prompt, optimized jointly through model queries. This approach targets vulnerabilities in multi-modal understanding, posing risks to applications like autonomous driving and content moderation. Existing attacks typically focus on single modalities or require white-box access. The paper is available on arXiv under ID 2605.26501.