Adversarially Robust Neural Architectures.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep Neural Networks (DNNs) are susceptible to adversarial attacks.
• Current defense strategies primarily focus on modifying network weights through robust training or regularization.
• The role of neural network architecture in adversarial robustness remains largely unexplored.

Purpose of the Study:

• To improve the adversarial robustness of DNNs from an architectural perspective.
• To investigate the relationship between adversarial robustness, Lipschitz constant, and architecture parameters.
• To develop a method for constraining architecture parameters to enhance robustness.

Main Methods:

• Explored the link between adversarial robustness, Lipschitz constant, and architecture parameters.
• Constrained architecture parameters to reduce the network's Lipschitz constant.
• Approximated the network's Lipschitz constant using a log-normal distribution related to architecture parameters.

Main Results:

• Demonstrated that constraining architecture parameters effectively reduces the Lipschitz constant, thereby improving adversarial robustness.
• The proposed method achieved superior performance compared to existing adversarially trained and human-designed models.
• Empirical validation showed best performance across various attacks and datasets.

Conclusions:

• Neural network architecture is a critical, yet underexplored, factor in adversarial robustness.
• Architectural modifications offer a promising new direction for defending DNNs against adversarial attacks.
• The proposed approach provides a novel and effective method for enhancing DNN security.