Digital thermal infrared detector attack via free velocity and rollback mutation

Existing black -box attack methods for infrared detectors often rely on heuristic techniques due to the unavailability of useful gradient information from the target detection model. However, existing heuristicbased attack methods suffer from the following two drawbacks. First, they are often prone to falling into local optima. Second, their convergence speed is particularly slow in the later stages of the optimization algorithm. To address these challenges, we propose a thermal infrared detector Attack (TID-Attack) based on free velocity and rollback mutation. This algorithm enables effective black -box digital adversarial attacks on infrared object detection models. Specifically, we first introduce a free velocity attack method in the particle swarm optimization algorithm. This method effectively balances the local and global search capabilities of particles during the optimization process, mitigating the risk of particles getting trapped in local optima. Additionally, we design a rollback mutation search strategy that allows particles trapped in local optima to bounce to new areas, farther away from their current positions, and then perform the optimization process again. These two modules make heuristic -based attack methods more robust and better stable. To evaluate the effectiveness of TID-Attack, we perform black -box attack tests on the YOLOv5 and YOLOv3 using three infrared detection datasets: FLIR-ADAS V2, CVC-09,(Daytime and Nighttime), and KAIST. Extensive experimental results demonstrate that our method achieves superior performance in terms of attack success rate and query times.