Adversarial Training With Anti-Adversaries.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Adversarial training improves deep neural network robustness but faces challenges in fairness, generalization, and robustness.
• Existing methods often use uniform perturbation directions and bounds, limiting their effectiveness.

Purpose of the Study:

• To theoretically and practically investigate the impact of diverse perturbation directions (adversarial and anti-adversarial) and bounds on model performance.
• To develop a more general adversarial training framework addressing fairness, robustness, and generalization trade-offs.

Main Methods:

• Theoretical analysis of adversarial training under a generalized perturbation scope with varied directions and bounds per sample.
• Introduction of a new learning objective combining adversaries and anti-adversaries with sample-specific bounds.
• Development of meta-learning and reinforcement learning-based frameworks to dynamically determine perturbation strategies.
• Explanation of the varied bounds strategy from a regularization perspective.

Main Results:

• Theoretical insights suggest combining varied adversarial and anti-adversarial perturbations improves fairness and the robustness-accuracy-fairness trade-off.
• The proposed frameworks effectively determine sample-specific perturbation directions and bounds.
• Extensive experiments confirm theoretical findings and the efficacy of the proposed methodology across various scenarios.

Conclusions:

• The proposed adversarial training approach with varied perturbation directions and bounds offers significant improvements in fairness, robustness, and generalization.
• This generalized framework provides a more effective strategy for training robust and fair deep learning models.