M-Patch: Adversarial Patch-based Physical Attack against Multi-modal Fusion Detectors

As reliance on multimodal fusion object detection models grows in critical tasks, ensuring their robustness and security is essential before deployment. However, research on adversarial attacks in multimodal target detection is still in its early stages, making it crucial for evaluating and improving system resilience. To address these challenges, we propose M-Patch, an adversarial patch-based attack method specifically designed for multimodal fusion object detection models. M-Patch can execute attacks in both digital and physical environments and is applicable to various multimodal detection models. In digital tests, M-Patch achieved success rates exceeding 60%, with a notable 97.6% success rate against the MFEIF+YOLOv3 multimodal detection model. Physical experiments further validated the patch's effectiveness, demonstrating strong attack results. More importantly, M-Patch reveals significant vulnerabilities in multimodal systems, offering a practical and efficient approach to enhancing their security.