Knowledge-Guided Adversarial Training (KGAT) for Infrared Object Detection Based on Thermal Radiation Modeling

The researchers propose Knowledge-Guided Adversarial Training (KGAT), which embeds physical knowledge of infrared thermal radiation into adversarial training to improve the robustness of infrared object detection. This method models inter-class thermal radiation relationships based on the ordering of grayscale values across classes and quantifies their stability, making prediction results conform to actual physical laws. Experiments on three infrared datasets and six mainstream infrared object detection models show that KGAT simultaneously improves clean-sample accuracy and robustness against adversarial attacks and common corruptions.