Potential auto-driving threat: Universal rain-removal attack

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Driving

Background:

• Severe weather challenges autonomous driving perception.
• Image rain-removal algorithms use neural networks to restore clear images.
• Existing algorithms lack robustness against adversarial attacks.

Purpose of the Study:

• To propose a universal rain-removal attack (URA) targeting image rain-removal algorithms.
• To investigate the vulnerability of neural networks in autonomous driving perception systems.
• To evaluate the effectiveness of URA in degrading image quality and scene restoration.

Main Methods:

• Developed a universal rain-removal attack (URA) using non-additive spatial perturbation.
• Applied URA to state-of-the-art image rain-removal algorithms.
• Quantified the impact of URA on scene restoration similarity and image quality.

Main Results:

• URA significantly diminishes scene restoration similarity and image quality.
• The attack reduced scene repair capability by 39.5%.
• Image generation quality was reduced by 26.4%.

Conclusions:

• URA is an effective method for detecting vulnerabilities in image rain-removal algorithms.
• The imperceptible and generic nature of URA suggests potential for real-world AI attacks.
• Enhanced robustness is crucial for AI perception systems in adverse weather conditions.