Effective and Efficient Adversarial Detection for Vision-Language Models via A Single Vector
This addresses a critical safety issue for VLMs by improving adversarial detection, though it is incremental as it builds on existing detection approaches with a new dataset and method.
The paper tackles the problem of adversarial image detection in Vision-Language Models (VLMs) by constructing a large-scale dataset (RADAR) and proposing a detection method (NEARSIDE) that uses a single vector from hidden states, achieving effective and efficient results with cross-model transferability as demonstrated on LLaVA and MiniGPT-4.
Visual Language Models (VLMs) are vulnerable to adversarial attacks, especially those from adversarial images, which is however under-explored in literature. To facilitate research on this critical safety problem, we first construct a new laRge-scale Adervsarial images dataset with Diverse hArmful Responses (RADAR), given that existing datasets are either small-scale or only contain limited types of harmful responses. With the new RADAR dataset, we further develop a novel and effective iN-time Embedding-based AdveRSarial Image DEtection (NEARSIDE) method, which exploits a single vector that distilled from the hidden states of VLMs, which we call the attacking direction, to achieve the detection of adversarial images against benign ones in the input. Extensive experiments with two victim VLMs, LLaVA and MiniGPT-4, well demonstrate the effectiveness, efficiency, and cross-model transferrability of our proposed method. Our code is available at https://github.com/mob-scu/RADAR-NEARSIDE