Sample selection of adversarial attacks against traffic signs.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Cybersecurity

Background:

• Accurate traffic sign recognition is vital for autonomous driving and traffic monitoring.
• Adversarial attacks test the security of autonomous driving systems and inform algorithm improvements.
• Evolving transportation infrastructure necessitates adaptive adversarial attack models for new traffic sign types.

Purpose of the Study:

• To develop a class incremental learning method for traffic sign adversarial attacks.
• To enable adversarial attack models to adapt to the addition of new traffic sign classes.
• To enhance the security testing and robustness of autonomous driving recognition algorithms.

Main Methods:

• Proposes a class incremental learning approach for traffic sign adversarial attacks.
• Utilizes Pinpoint Region Probability Estimation Network (PRPEN) to identify high-probability attack regions in existing samples.
• Constructs a replay sample set prioritizing older samples with smaller high-probability pixel concentration areas for incremental learning.

Main Results:

• The proposed method selects more representative samples compared to existing methods.
• Enables more effective training of PRPEN for generating accurate probability maps.
• Demonstrates improved adversarial attack generation on traffic signs within an incremental learning framework.

Conclusions:

• The developed class incremental learning method enhances the adaptability of adversarial attack models to new traffic sign classes.
• Prioritizing specific samples improves the efficiency and effectiveness of training adversarial attack models.
• Contributes to more robust security evaluations and improvements for autonomous driving systems.