Variational Adversarial Defense: A Bayes Perspective for Adversarial Training.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Existing adversarial defense methods lack theoretical guarantees, leading to issues like overfitting and gradient masking.
• Point-wise adversarial sampling provides insufficient support, hindering the formation of robust decision boundaries.

Purpose of the Study:

• To theoretically analyze the relationship between robust accuracy and training set complexity in adversarial training.
• To propose a novel, distribution-wise adversarial training scheme for enhanced defense capabilities.

Main Methods:

• Theoretical analysis of robust accuracy and training set complexity.
• Development of Variational Adversarial Defense (VAD) for distribution-wise adversarial sample generation.
• Taylor expansion technique for analyzing the proposed method's interpretability.

Main Results:

• Variational Adversarial Defense upgrades defense from point-wise to distribution-wise sampling.
• The method enlarges the support region for adversarial data, improving training robustness.
• Augmenting the training set with a larger support region enhances decision boundary smoothness.

Conclusions:

• Variational Adversarial Defense offers a theoretically grounded approach to enhance adversarial robustness.
• The distribution-wise strategy effectively addresses limitations of point-wise sampling.
• The method provides a more interpretable and robust defense against adversarial attacks.