Physical Adversarial Attacks for Infrared Object Detection

Utilizing the physical adversarial methods to attack infrared objects is on the rise. However, existing physical infrared attacks are complex to implement in real world due to professional and cumbersome making processes. And these attack methods only target infrared pedestrian objects. In this work, we propose an alternative physical method to attack infrared pedestrian and infrared car objects, named Sensitive Region Patches, with low-cost and easy making process. Given that the imaging mechanism of infrared imager by capturing the thermal radiation of the obj ect, Sensitive Region Patches perform attacks by attaching the aerogel patch, a thermal insulation material to the Sensitive Region (SR) of the object to alter its thermal distribution. Therefore, we propose a consensus selection strategy to find out the Sensitive Region (SR), and cover such SR in the obj ect with the aerogel patch cropped based on SR to achieve the physical attack. We conduct experimental measurements both in infrared car detection and infrared pedestrian detection to verify our Sensitive Region Patches. Extensive experiments show that our method achieves over 70% Attack Success Rate (ASR) under various physical conditions, including shooting angles, distances, postures, and scenes, and Sensitive Region Patches can be constructed physically in just 0.25 hours, demonstrating the effectiveness and simplicity of our method.