DCT-based Fast Localization and Elimination of Adversarial Patches*

As autonomous driving technology advances, the role of image recognition models in autonomous driving systems has become increasingly critical, particularly object detection models, which are vital for environmental perception and safety decision-making. However, recent studies have shown that adversarial attacks can compromise the accuracy of object detection models without being perceptible to the human eye, posing a substantial threat to the safety of autonomous driving systems. To tackle these challenges, a fast and accurate defense framework without additional training is proposed that directly eliminates adversarial patches by employing a scale-adaptive discrete cosine transform (DCT) method to detect the anomalous frequency components introduced by adversarial patches. The framework then restores the original semantic texture of the image through an image inpainting method. Our suggested defense system seeks effective protection while minimizing computational demands to accommodate potential real-world scenarios. We evaluated our framework on multiple datasets of different types, and the results demonstrate that our method significantly improves the robustness of object detection models against adversarial attacks.