Improving Adversarial Robustness of Deep Neural Networks via Adaptive Margin Evolution

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep Neural Networks (DNNs) are vulnerable to adversarial attacks.
• Adversarial training enhances DNN robustness but often requires extensive hyperparameter tuning.
• Existing methods are sensitive to hyperparameters like the training noise upper bound, limiting practical applications.

Purpose of the Study:

• To introduce Adaptive Margin Evolution (AME), a novel, hyperparameter-free adversarial training method.
• To improve the accessibility and applicability of adversarial training techniques.
• To enhance Deep Neural Network (DNN) robustness against adversarial noises.

Main Methods:

• Developed Adaptive Margin Evolution (AME), a method that optimizes adversarial training sample placement.
• AME utilizes self-adaptive and gradient-aware step sizes to expand the exploration range.
• Evaluated AME against seven other adversarial training methods on CIFAR10, SVHN, and Tiny ImageNet datasets using AutoAttack.

Main Results:

• AME demonstrated superior overall performance across all three benchmark datasets.
• AME achieved the best performance at all noise levels on the challenging Tiny ImageNet dataset.
• The proposed method significantly improves DNN robustness without requiring user-defined hyperparameters.

Conclusions:

• Adaptive Margin Evolution (AME) offers a robust and user-friendly approach to adversarial training.
• AME's hyperparameter-free nature facilitates the adoption of adversarial robustness techniques in diverse application domains.
• This research paves the way for more widespread use of adversarial training in real-world scenarios.