Adversarial rain attack and defensive deraining for DNN perception.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) are vulnerable to adverse weather conditions like rain, impacting perception systems.
• Synthesizing realistic rainy images for comprehensive testing and training is challenging.
• Existing methods struggle to capture the diverse range of real-world rain scenarios.

Purpose of the Study:

• To investigate the risks posed by rain to DNN-based perception systems.
• To develop a method for synthesizing diverse and realistic rainy images.
• To enhance the robustness of DNNs against rain and improve deraining models.

Main Methods:

• Combined adversarial attack techniques with rainy image synthesis.
• Developed a factor-aware rain generator simulating camera exposure and learnable rain factors.
• Conducted adversarial rain attacks on image classification and object detection tasks.
• Proposed an adversarial rain augmentation strategy for defensive deraining models.

Main Results:

• Synthesized rainy images exhibit realistic appearances and strong adversarial capabilities against DNNs.
• The adversarial rain augmentation effectively enhanced the performance of deraining models.
• Demonstrated significant improvements in DNN robustness and deraining model performance on large-scale datasets.

Conclusions:

• The proposed adversarial rain synthesis effectively reveals threats to DNNs and enhances their robustness.
• The adversarial rain augmentation strategy provides a foundation for developing more resilient perception systems.
• This work paves the way for future research in rain-robust computer vision.