Strength-Adaptive Adversarial Training.

Area of Science:

• Machine Learning
• Computer Vision
• Deep Learning

Background:

• Conventional adversarial training (AT) uses a fixed perturbation budget.
• This fixed budget leads to robustness disparities and overfitting in models with varying capacities.
• Attack strength does not adapt to the model's evolving robustness during training.

Purpose of the Study:

• To introduce Strength-Adaptive Adversarial Training (SAAT) for enhanced model robustness.
• To address limitations of fixed perturbation budgets in conventional AT.
• To mitigate robust overfitting and control accuracy disparities.

Main Methods:

• SAAT incorporates an adversarial-loss constraint to guide adversarial data generation.
• The perturbation budget dynamically adapts based on the current training state.
• Attack strength is precisely regulated through adversarial loss.

Main Results:

• SAAT effectively mitigates robust overfitting.
• SAAT enables precise control over the robustness disparity between natural and adversarial accuracies.
• Experiments show SAAT substantially improves adversarial robustness compared to standard AT.

Conclusions:

• SAAT offers a superior approach to adversarial training.
• Dynamic adaptation of perturbation budgets enhances model resilience.
• SAAT provides better control over model robustness and accuracy trade-offs.