MC-SNN: Multicenter Stochastic Neural Network for Adversarially Robust Learning

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep neural networks (DNNs) often struggle with adversarial robustness due to single-center learning modes that fail to capture complex data distributions.
• Existing methods like standard training (ST) and adversarial training (AT) can degrade robustness by misclassifying samples near decision boundaries.

Purpose of the Study:

• To propose a novel multicenter learning method for enhancing the adversarial robustness of DNNs.
• To address the limitations of single-center learning in fitting complex, multi-modal data distributions in latent space.

Main Methods:

• Introduced a multicenter stochastic neural network (MC-SNN) that leverages feature uncertainty learning to induce multiple class centers in latent space.
• Developed MC-SNN-AT by integrating four adversarial training strategies to defend against diverse attacks.
• Utilized benchmark tests to evaluate the proposed methods' performance.

Main Results:

• Both MC-SNN and MC-SNN-AT demonstrated state-of-the-art adversarial robustness.
• MC-SNN achieved robustness enhancement with a training cost approximately one-tenth of vanilla adversarial training.
• The multicenter approach allows for more delicate data fitting, improving classification accuracy and robustness.

Conclusions:

• The proposed multicenter learning approach significantly enhances adversarial robustness in DNNs.
• MC-SNN offers a computationally efficient alternative to traditional adversarial training methods.
• This method provides a promising direction for developing more reliable and secure artificial intelligence systems.