Perturbation diversity certificates robust generalization.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Adversarial training is effective against adversarial attacks on deep neural networks but suffers from overfitting and poor generalization.
• Conventional methods generate biased adversarial examples, leading to inhomogeneous data distributions and limited robustness.

Purpose of the Study:

• To propose a new method for generating diverse adversarial examples to improve robust generalization.
• To mitigate the limitations of conventional supervised adversarial training.

Main Methods:

• Generate adversarial examples from a perturbation diversity perspective, ensuring samples are both adversarial and diverse.
• Provide theoretical and empirical analysis to support the proposed method.
• Prove that promoting perturbation diversity improves the robust generalization bound.

Main Results:

• The proposed method generates diverse and adversarial samples, leading to a more homogeneous data distribution.
• Theoretical analysis confirms that perturbation diversity enhances robust generalization bounds.
• Extensive experiments on CIFAR-10, CIFAR-100, and SVHN datasets demonstrate superior performance over state-of-the-art methods like PGD and Feature Scattering.

Conclusions:

• Generating diverse adversarial examples is key to achieving robust generalization in deep neural networks.
• The proposed perturbation diversity approach offers significant improvements in robustness and generalization compared to existing methods.
• This work provides a strong theoretical and empirical foundation for future research in robust adversarial training.