Exploring Transferability Optimization Methods for Physical Adversarial Attacks in Object Detection in Real-World Scenarios

The advent of deep neural networks (DNNs) and convolutional neural networks (CNNs) has revolutionized object detection by eliminating reliance on hand-crafted feature engineering and achieving near-human-level accuracy in specific domains. These advancements have enabled safety-critical applications such as autonomous driving, robotics, and video surveillance. Contemporary object detection methods predominantly employ DNNs as backbone architectures for hierarchical feature extraction, object classification, and spatial localization. However, their well-documented vulnerability to adversarial attacks raises significant concerns regarding security and reliability, particularly in physical-world deployments where environmental variables such as spatial deformations and illumination variations introduce additional constraints. Notably, adversarial examples demonstrate transferability across models, enabling cross-architecture generalization and empowering black-box attacks without requiring prior knowledge of the target model's architecture. In this study, we present a systematic and comprehensive survey of methodological advancements in enhancing the transferability of physical adversarial attacks for object detection systems. Our analysis unfolds through three principal dimensions: First, we establish a theoretical framework for transferability analysis and propose evaluation criteria specifically tailored to physical adversarial attacks. Second, we conduct a chronological taxonomy of seminal contributions in this domain, structuring our review through technical lineages to elucidate evolutionary trajectories. Finally, we identify persisting technical challenges and propose actionable research directions, thereby establishing foundational insights for understanding the operational vulnerabilities of deep learning systems in real-world scenarios. This survey aims to provide researchers, practitioners, and policy-makers with a structured knowledge framework to accelerate progress in physical adversarial research and strengthen security protocols for deep learning-based object detection systems in mission-critical applications.