Hiding from infrared detectors in real world with adversarial clothes

Thermal infrared detection is widely used in many scenarios including fast body temperature monitoring, safety monitoring and autopilot, however, its safety research has not attracted sufficient attention. We proposed the adversarial clothing to test the safety of infrared detection, which could hide from infrared detectors in the real world. The adversarial clothing uses flexible carbon fiber heaters as the basic elements. We optimized the patterns formed by different heaters based on the adversarial example technique. The optimized pattern lowered the average precision (AP) of YOLOv3 by 66.33%, while the random pattern lowered the AP by only 31.33% in the digital world. We then manufactured the adversarial clothing and tested the safety of infrared detectors in the physical world. The adversarial clothing lowered the AP of YOLOv3 by 43.95%, while the clothing with randomly placed heaters lowered the AP of YOLOv3 by only 19.21%. With ensemble attack techniques, our attack method had good transferability to unseen CNN models. We tested five typical defense methods but achieved limited success. These results indicate that current thermal infrared detectors are not robust. Graphical Abstract