针对目标检测的可迁移性对抗补丁生成方法OA

With the widespread deployment of object detection models in real-world applications,their security is-sues have increasingly become a research focus.Adversarial attack techniques,by carefully designing adversarial patches,can effectively induce models to produce erroneous predictions,thereby revealing the inherent vulnerabilit-ies of deep neural networks in the decision-making process.To enhance the transferability of adversarial patches across different detectors,most existing methods rely on static weight fusion strategies for joint optimization.However,such approaches struggle to fully reconcile the discrepancies in vulnerability distributions and optimiza-tion dynamics among detectors,leading to imbalanced attack effectiveness across models and significantly limiting the transferability.To address this challenge,this paper proposes a transferable adversarial patch generation frame-work based on a multi-task dynamic reweighting mechanism.The framework introduces a global correction factor and a local correction factor,which dynamically adjust task weights from two perspectives:The overall optimiza-tion progress among tasks and the fine-grained convergence behavior of individual tasks.This design enables better coordination and improved robustness during multi-model joint optimization.Extensive experiments in both the di-gital and physical domains demonstrate that the proposed method significantly enhances the adversarial transferab-ility of patches across various object detectors and achieves strong attack performance in deployments under real-world physical domain.