Robust Adversarial Patch for Object Detection Using Self-Similarity for Multiscale Attacks

Research on adversarial examples for deep learning models has gained significant attention, particularly in object detection, where dynamic relative positioning between the object and the detector poses unique challenges. Existing attacks remain limited in handling real-world scenarios, especially regarding distance variations that induce blurring, reducing patch effectiveness. Additionally, as the target carrying the adversarial patch moves closer, incomplete patch capture may further weaken its impact. To address these issues, we propose a systematic approach to designing robust adversarial patches for object detectors. Our method leverages self-similarity to ensure attack effectiveness across scales and incorporates a blur simulation module to enhance resilience against distance-induced distortions. Furthermore, we introduce two key geometric optimization metrics, Local Patch Shift (LPS) and Embedded Boundary Gap (EBG), to analyze the impact of local perturbations on global patch optimization. Experimental results demonstrate that our adversarial patch achieves a high attack success rate on several object detection models in digital environments. In addition, we perform evaluations in simulated and physical environments, validating the robustness of the patch under varying lighting, motion, and detection angles. The findings highlight the effectiveness of our approach in both simulated and real-world settings, providing valuable insights into improving adversarial patch design for practical applications.