AEDD: Anomaly Edge Detection Defense for Visual Recognition in Autonomous Vehicle Systems

Lihao Wang, Junwei Zhou, Dongdong Zhao, Dongqing Liao, Mengjie Wang, Jianwen Xiang

Abstract: Visual recognition algorithms based on deep neural network (DNN) have been widely used in the design of automatic driving to recognize traffic sign images. However, there exists adversarial patches which are essentially the anormal image block that can be locally observed but not noticed by humans. And these visual recognition algorithms often suffer from the effect of adversarial patches, due to these patches can change the algorithms recognition result of the images. To solve the above issues, this work proposes the anomaly edge detection and image inpainting defense (AEDD) for visual recognition. This framework uses anomaly location to obtain the anomaly area, uses edge detection to get an accurate edge of the anomaly area, finally uses image inpainting to repair this area. We also combine two attack algorithms with three patch sizes, and generate six types of adversarial patches on the GTSRB dataset. We have demonstrated the effectiveness of our approach, resulting in an average 6.6% increase in defense accuracy compared to the state-of-the-art methods. Our code is available at https://github.com/drtt438/AEDD for the purpose of reproducibility.