Toward Evaluating the Robustness of Deep Learning Based Rain Removal Algorithm in Autonomous Driving

Abstract

Autonomous driving systems have been widely adopted by automobile manufacturers, ushering in a new era of intelligent transportation. While adverse weather conditions continue to pose a significant challenge to its commercial application, as they can impact sensor data, degrade the quality of image transmission, and pose safety risks. Using neural network models to remove rain has shown significant promise in addressing this problem. The learning-based rain-removal algorithm discovers the deep connection between rainy pictures and non-rainy pictures by mining the information on raindrops and rain patterns. Nevertheless, the robustness of these rain removal algorithms was not considered, which poses a threat to autonomous vehicles. In this paper, we propose an optimized CW adversarial sample attack to explore the robustness of the rain removal algorithm. In our attacks, we generate a perturbation index of structural similarity that is difficult to detect through human vision and image pixel analysis, causing the similarity and image quality of the restored scene to be significantly degraded. To validate the realistic attack potential of the proposed method, a pre-trained State-of-the-art rain removal attack algorithm, RainCCN, is used as a potential victim of the proposed attack method. We demonstrate the effectiveness of our approach against a state-of-the-art rain removal algorithm, RainCCN, and show that we can reduce PSNR by 39.5 and SSIM by 26.4.