Design of a Global Adversarial Attack Scheme for YOLOv8

Deep learning-based object detection technology has rapidly advanced, with state-of-the-art detectors like YOLOv8 being widely utilized in fields such as autonomous driving. Research indicates that these object detectors are highly susceptible to adversarial attacks. In response to this vulnerability, we propose an imperceptible global perturbation attack method specifically designed for the advanced YOLOv8 detector. This method builds on the Gradient Iterative Attack by incorporating the Adam optimization algorithm, effectively addressing the problem of suboptimal local minima during gradient iterations and thereby enhancing the attack's performance. Experimental results on the VOC dataset show that our proposed method can reduce the detection accuracy of the YOLOv8 detector to 2.6%, with a 1.5% improvement in attack success rate compared to the baseline TOG (Targeted Adversarial Objectness Gradient Attacks) method. These results clearly demonstrate the efficiency and effectiveness of the proposed attack strategy.