Enhancing DNN Adversarial Robustness via Dual Stochasticity and Geometric Normalization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep neural networks (DNNs) are powerful but vulnerable to adversarial attacks, limiting their use in safety-critical systems.
• Existing stochastic defenses often use fixed noise and ignore decision space geometry, leading to suboptimal robustness.

Purpose of the Study:

• To introduce Dual Stochasticity and Geometric Normalization (DSGN), a novel framework to enhance adversarial robustness of DNNs.
• To address limitations of current defenses by incorporating input-dependent noise and geometric stability.

Main Methods:

• DSGN uses learnable, input-dependent Gaussian noise in feature representations and classifier weights for dual-path stochastic modeling.
• L2 normalization projects noisy components onto a unit hypersphere, stabilizing decision geometry and promoting margin separation.
• This approach captures multi-level predictive uncertainty and enhances decision consistency.

Main Results:

• DSGN demonstrated significant improvements in robust accuracy against PGD (1-6%) and AutoAttack (1-17%) on benchmark datasets and CNNs.
• The framework effectively stabilizes decision boundaries and representation geometry.
• High clean accuracy was maintained alongside enhanced adversarial robustness.

Conclusions:

• DSGN offers a promising approach to improve the adversarial robustness of deep neural networks.
• The combination of dual stochasticity and geometric normalization effectively enhances security for safety-critical applications.
• This method provides a more stable and robust defense mechanism compared to existing techniques.