Black-box physical adversarial stripes for hiding from infrared detectors at multiple views.

Thermal infrared detectors are widely employed in applications such as autonomous driving and night-time surveillance, making it essential to investigate their robustness against adversarial attacks. However, many existing works primarily focus on the attack effectiveness, while showing limitations in multi-view capability, stealthiness, and applicability under realistic black-box conditions. To hide from the thermal infrared detectors at multiple views stealthily, we propose Adversarial Stripes (AdvStripes), a novel black-box physical attack using periodic zigzag-shaped stripe patches. Our method aims to enhance multi-view effectiveness by improving the visual stability of adversarial perturbations. To achieve this, we design adversarial stripes attached circumferentially to pedestrians' bodies, maintaining a consistent shape and appearance from any perspective. To address stealthiness, we leverage low-temperature materials to craft stripes which are still effective underneath the clothing. An algorithm based on particle swarm optimization is employed to optimize the stripes' parameters, due to its suitability for black-box attacks. Moreover, the proposed stripe design is simple to implement in physical settings. Extensive experiments are conducted to demonstrate the effectiveness, stealthiness, and robustness of our method. The comparison with baseline methods presents the superiority of the proposed AdvStripes attack. Ablation studies, transferability tests, and defense strategies experiments are exhibited to gain deeper insights into our attack's mechanisms and potential impact.