Adversarial Robustness Via Fisher-Rao Regularization.

Area of Science:

• Machine Learning
• Computer Vision
• Deep Learning

Background:

• Neural networks exhibit brittleness, making them vulnerable to adversarial attacks.
• Adversarial robustness is a critical area of research in machine learning.

Purpose of the Study:

• To propose an information-geometric formulation for adversarial defense.
• To introduce Fire, a Fisher-Rao regularization method, to improve model robustness.

Main Methods:

• Developed an information-geometric formulation for adversarial defense.
• Introduced Fire, a Fisher-Rao regularization for categorical cross-entropy loss.
• Derived explicit characterization of Fisher-Rao Distance (FRD) for binary and multiclass cases.

Main Results:

• Fire reaches Pareto-optimal points in the accuracy-robustness region, outperforming other methods.
• Empirical evaluation shows up to 1% simultaneous improvement in clean and robust performances.
• Achieved a 20% reduction in training time compared to state-of-the-art methods.

Conclusions:

• Fire offers a novel and effective approach to enhance adversarial robustness in machine learning.
• The method provides a theoretical foundation in information geometry for adversarial defense.
• Fire demonstrates practical benefits in performance improvement and training efficiency.