THE SEMANTIC-FOCUS ADVERSARIAL EXAMPLES ON ANCHOR-FREE MODEL FOR OBJECT DETECTION

Deep neural networks are vulnerable to adversarial examples. Currently, there is relatively little research on adversarial examples for anchor-free object detectors, making such models more susceptible to adversarial attacks. To address this issue, we adopted a method of generating adversarial examples on anchor-free object detection models. This framework rapidly collected gradients based on the identified classes, which was more efficient than methods that generated perturbations based on individual candidate boxes. Meanwhile, a method for extracting semantic information masks was proposed, enabling the adversarial perturbations to be concentrated only in the semantically rich regions of the image, resulting in sparser and more focused perturbations. The results on two datasets demonstrate that this method achieves state-of-the-art performance in both white-box and black-box experiments, providing support for the improvement and optimization of the robustness of such networks.