Reproduce “Adversarial Robustness for Visual Grounding of Multimodal Large Language Models” on OpenResearch

About this paper

Kuofeng Gao, Yang Bai, Jiawang Bai, Yong Yang, Shu-Tao Xia

A study conducted by Tsinghua University and Tencent Security Platform revealed the adversarial vulnerability of visual grounding capabilities in Multi-modal Large Language Models. The research introduced three adversarial attack paradigms, demonstrating a substantial reduction in correct object localization for untargeted attacks and successful redirection to incorrect bounding boxes for targeted attacks across benchmark datasets.

Key insights

• MLLMs' visual grounding capabilities are susceptible to imperceptible adversarial perturbations in input images, leading to mislocalization or incorrect object identification.
• Image embedding attacks, which disrupt foundational visual features, proved more effective in untargeted scenarios than textual bounding box attacks, which target the auto-regressive text generation process.
• Exclusive targeted attacks, which aim for a fixed, universal incorrect bounding box, achieved higher success rates than permuted targeted attacks, which require shifting predictions to the locations of other specific objects.

Summary by alphaXiv.