Improving Fast Adversarial Training With Prior-Guided Knowledge.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Fast adversarial training (FAT) enhances model robustness against white-box attacks but is prone to catastrophic overfitting.
• Existing FAT variants to mitigate overfitting often incur significant training time penalties.

Purpose of the Study:

• Investigate the link between adversarial example quality and catastrophic overfitting in FAT.
• Propose a novel method to prevent overfitting and improve adversarial robustness without increasing training time.

Main Methods:

• Compared standard adversarial training and FAT to understand overfitting dynamics.
• Developed a positive prior-guided adversarial initialization using historical high-quality perturbations.
• Introduced prior-guided regularization for loss function smoothness and a prior-guided ensemble FAT method.

Main Results:

• Established that catastrophic overfitting correlates with declining adversarial example attack success rates.
• The proposed FGSM-PGK method effectively prevents overfitting by enhancing adversarial example quality.
• Demonstrated superior adversarial robustness on four datasets compared to existing methods.

Conclusions:

• Prior knowledge, specifically high-quality adversarial perturbations and model weights, can prevent FAT overfitting.
• The proposed FGSM-PGK method offers an efficient solution for improving adversarial robustness in white-box scenarios.