Between-Class Adversarial Training for Improving Adversarial Robustness of Image Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep neural networks (DNNs) are vulnerable to adversarial attacks.
• Adversarial training (AT) is the primary defense but suffers from a trade-off between robustness and standard accuracy.
• Improving this trade-off is crucial for reliable DNN deployment.

Purpose of the Study:

• To propose a novel defense algorithm, Between-Class Adversarial Training (BCAT), to enhance DNN robustness and generalization.
• To address the accuracy trade-off inherent in standard adversarial training.
• To introduce BCAT+ with an improved mixing strategy for even greater efficacy.

Main Methods:

• BCAT combines Between-Class learning (BC-learning) with standard AT.
• The method trains models using mixed adversarial examples from different classes.
• BCAT+ utilizes a more potent mixing technique for feature distribution regularization.

Main Results:

• BCAT and BCAT+ effectively regularize adversarial example feature distributions, increasing between-class distances.
• The proposed algorithms demonstrated superior global robustness generalization compared to state-of-the-art methods.
• Evaluations on CIFAR-10, CIFAR-100, and SVHN datasets showed significant improvements under various attacks.

Conclusions:

• BCAT and BCAT+ offer improved robustness and generalization for DNNs against adversarial attacks.
• The methods successfully mitigate the accuracy trade-off without introducing new hyperparameters.
• These algorithms represent a significant advancement in defending DNNs in real-world applications.