A Survey of Physical Adversarial Attacks Against Infrared Target Detection

The existence of adversarial examples revealed the underlying vulnerabilities of deep learning models. In the digital domain, meticulously designed subtle perturbations could degrade model performance or even completely cause it to lose its intended functionality. In real-world physical scenarios, adversarial examples also demonstrated potent destructive power and substantial risks. This survey explored the issue of physical adversarial attacks in the crucial field of infrared object detection. Firstly, it provided a clear and in-depth explanation of the concept of physical adversarial attacks in the field of infrared target detection, distinguishing them from traditional digital adversarial attacks, and precisely locating their unique roles and manifestations in real-world scenarios. Secondly, focusing on the diverse detection tasks and targets in infrared scenarios, this review comprehensively and thoroughly examined the emerging physical adversarial attack methods targeting infrared object detection network models, from the perspectives of different detection systems such as pedestrian, vehicle, and remote sensing data detection. Finally, based on the then-current research achievements and technological development trends, it offered a positive outlook for the future of this field.