Text-Guided Gradient Refinement: Resolving Multimodal Gradient Conflicts to Boost Adversarial Attacks on Vision-Language Models

Vision-Language Models (VLMs) have advanced multimodal understanding, yet they remain susceptible to adversarial attacks. Among various strategies, transfer-based attacks are notably effective, especially in black-box scenarios. The dominant approach within this paradigm leverages generative models to create image targets from text, consistently outperforming text-only methods. However, this approach suffers from a fundamental limitation: generative models introduce visual features irrelevant or even detrimental to textual semantics, misguiding optimization. To investigate this limitation, we conduct comprehensive analysis revealing two critical findings. First, optimal attack directions lie in synergistic spaces between image and text gradients, demonstrating that text provides complementary information. Second, widespread gradient conflicts occur when combining modalities, where image-target gradients oppose text-target directions. This conflict provides direct evidence that extraneous visual information actively harms optimization, driving it away from intended textual objectives. Based on these insights, we propose Text-Guided Gradient Refinement (TGGR), a novel framework that employs a conflict-aware projection mechanism to resolve this conflict. TGGR preserves the beneficial characteristics of image targets by decomposing the image gradient and surgically removing components that oppose the textual guidance. Extensive experiments on models such as LLaVA and GPT-4o demonstrate that TGGR substantially improves attack success rates. Specifically, on GPT-4o, TGGR yields an improvement of up to 14% over state-of-the-art methods, achieving 96% attack success rate. Our work offers a principled framework for developing more synergistic and effective adversarial strategies against VLMs.