Sparse Adversarial Patch Attack and Robustness Evaluation Algorithm for Vision-Language Models
Visual language models (VLMs) have demonstrated outstanding performance in high-value domains such as autonomous driving, unmanned system navigation, and intelligent question-answering; however, the security of their cross-modal alignment mechanisms has not yet been fully verified. Existing visual adversarial patch attacks typically rely on continuous, dense pixel perturbations, which are easily detected and blocked by anomaly detection systems in practical engineering applications. This paper proposes a novel sparse adversarial patch attack algorithm (Sparse Patch Attack, SPA), which successfully misleads the text generation results of VLMs by generating highly dispersed discrete pixel perturbations in non-salient regions of the image. To achieve this, we introduce a differentiable L0-norm approximation and a cross-attention masking mechanism to minimize the number of modified pixels. Furthermore, addressing the characteristics of large-scale model open-ended text generation, we construct a multi-dimensional robustness evaluation framework covering semantic deviation, target achievement rate, and visual concealment. Preliminary experiments on mainstream visual-language models (such as LLaVA and BLIP-2) demonstrate that the SPA algorithm can achieve high success rates in targeted cross-modal attacks with an extremely low pixel modification rate (<1%). This study reveals a novel security vulnerability in visual-language models within complex real-world environments and provides a quantitative evaluation benchmark for future defense mechanisms in multimodal models.