CodeMosaic Patch: Physical Adversarial Attacks Against Infrared Aerial Object Detectors

Hangwei He, Libing Wu, Enshu Wang, Yizhou Wang, Yu Zhao

Abstract: In recent years, drones illegally spy on military bases and steal military secrets, especially in dark scenes. In the face of illegal detection by drones, physical infrared adversarial patches can play a defensive role. Existing methods are either hard to implement physically or lose effectiveness over time. They also overlook that the target may be positioned at various angles. To solve this problem, we design the CodeMosaic Patch (CMP) method for attacks on infrared remote sensing object detectors, which combines remote sensing attacks and infrared attacks to hide non-live targets like vehicles. The CMP utilizes differential evolutionary algorithms to train the patch’s structure, texture, size, position, and rotation angle. We use aluminum foil as the reproduction material for infrared patches, successfully concealing the target object from infrared detectors in the real world. Experimental results show that the average digital and physical domain attack success rates (ASR) are 84.2% and 81.02%, respectively. Notably, our CMP can transition from infrared to visible light, achieving a 92.42% ASR and demonstrating cross-modality generalization. This represents the first instance of a patch with cross-modality attack performance being proposed.