Grid-Based Adversarial Patch Placement Optimization for Enhanced Object Detection Attacks*

As computer vision tasks improve capabilities across numerous applications, various attacks targeting these tasks have emerged. Among them, adversarial patches have gained significant attention for their ability to disrupt vision models in the physical world. To enhance the attack performance of patches, various methods have been proposed, focusing on factors such as size and location. While existing patch placement optimization algorithms improve attack effectiveness, they often rely on object size, limiting their generalized attack performance. In this paper, we proposed a grid-based adversarial patch placement optimization algorithm designed to enhance attack performance by maximizing disruption across detection grid cells. To evaluate its effectiveness, we compare the attack success rate (ASR) with the original algorithm and other algorithms by using the PASCAL VOC dataset. Under the same patch conditions, our algorithm achieves a 9.8% increase in ASR compared to the original algorithm. Furthermore, when integrated with explainable AI (XAI)-based placement, it achieves a 15.5% improvement over the original algorithm, demonstrating its effectiveness in enhancing adversarial patch attacks.