Simpler Certified Radius Maximization by Propagating Covariances.

Area of Science:

• Machine Learning
• Computer Vision
• Robustness in AI

Background:

• Adversarial training enhances model robustness by maximizing certified radius.
• Current methods rely on Monte Carlo sampling, creating a computational bottleneck.
• The certified radius defines a stable prediction neighborhood around training samples.

Purpose of the Study:

• To investigate mitigating the sampling bottleneck in adversarial training.
• To explore direct propagation of smoothed distribution covariance matrices through neural networks.
• To develop an algorithm for maximizing certified radius with improved efficiency.

Main Methods:

• Developing network adjustments and accounting for distributional moment transformations.
• Implementing an algorithm to directly propagate covariance matrices.
• Evaluating the approach on Cifar-10, ImageNet, and Places365 datasets.

Main Results:

• Achieved certified radius maximization on benchmark datasets.
• Demonstrated runtime savings for moderately deep networks.
• Observed a slight trade-off in overall accuracy.

Conclusions:

• Direct covariance propagation offers a viable alternative to Monte Carlo sampling for adversarial training.
• The proposed method provides efficiency gains with manageable accuracy compromises.
• Further experiments detail the practical applicability and limitations of the simplifications.