Provable Unrestricted Adversarial Training Without Compromise With Generalizability.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Adversarial training (AT) is crucial for defending against adversarial attacks.
• Existing AT methods struggle with unrestricted adversarial examples (UAEs) and often sacrifice generalizability for robustness.

Purpose of the Study:

• To propose a novel adversarial training approach, Provable Unrestricted Adversarial Training (PUAT).
• To enhance adversarial robustness against both UAEs and restricted adversarial examples (RAEs).
• To improve standard generalizability alongside adversarial robustness.

Main Methods:

• PUAT views UAEs as perturbed unobserved examples and addresses the distribution separation issue in AT.
• Utilizes partially labeled data and a novel augmented triple-Generative Adversarial Network (GAN) for effective UAE generation and natural data distribution capture.
• Integrates the target classifier's supervised loss into adversarial loss for distribution alignment.

Main Results:

• PUAT provides comprehensive adversarial robustness against both UAE and RAE.
• The method simultaneously improves the standard generalizability of the target classifier.
• Theoretical analysis and experiments confirm PUAT's superiority on benchmarks.

Conclusions:

• PUAT offers a promising solution to the limitations of existing adversarial training methods.
• The approach effectively balances adversarial robustness and standard generalizability.
• PUAT advances the field of robust machine learning and adversarial defense.