On the Real-World Adversarial Robustness of Real-Time Semantic Segmentation Models for Autonomous Driving.

Area of Science:

• Computer Vision
• Deep Learning Security

Background:

• Real-world adversarial examples (RWAEs), often as patches, pose significant risks to deep learning models in safety-critical applications like autonomous driving.
• Semantic segmentation (SS) models are particularly vulnerable to these attacks, impacting visual perception systems.

Purpose of the Study:

• To conduct a comprehensive evaluation of the robustness of semantic segmentation models against various adversarial patch types (digital, simulated, physical).
• To introduce novel methods for enhancing adversarial attacks and improving patch detection for SS models.

Main Methods:

• Developed a novel loss function to increase the effectiveness of pixel misclassification by attackers.
• Introduced an improved attack strategy to enhance the expectation over transformation (EOT) method for patch placement.
• Extended and optimized a state-of-the-art adversarial patch detection method for SS models, achieving real-time performance.

Main Results:

• Adversarial patch attacks, both digital and real-world, demonstrably affect SS model performance.
• The impact of these attacks is frequently localized around the patch's position in the image.
• The proposed detection method demonstrates real-time capabilities in real-world scenarios.

Conclusions:

• While adversarial patches present a visible threat, their localized impact suggests potential vulnerabilities in the spatial robustness of real-time SS models.
• Further research is needed to understand and mitigate the spatial effects of adversarial attacks on autonomous driving perception systems.