Weight-Covariance Alignment for Adversarially Robust Neural Networks
This provides a computationally efficient solution for improving adversarial robustness in neural networks, addressing a critical security issue in AI systems.
The paper tackles the problem of adversarial robustness in neural networks by proposing a stochastic neural network that learns anisotropic noise distributions to optimize a theoretical bound on robustness, achieving state-of-the-art performance without adversarial training on multiple benchmarks and architectures.
Stochastic Neural Networks (SNNs) that inject noise into their hidden layers have recently been shown to achieve strong robustness against adversarial attacks. However, existing SNNs are usually heuristically motivated, and often rely on adversarial training, which is computationally costly. We propose a new SNN that achieves state-of-the-art performance without relying on adversarial training, and enjoys solid theoretical justification. Specifically, while existing SNNs inject learned or hand-tuned isotropic noise, our SNN learns an anisotropic noise distribution to optimize a learning-theoretic bound on adversarial robustness. We evaluate our method on a number of popular benchmarks, show that it can be applied to different architectures, and that it provides robustness to a variety of white-box and black-box attacks, while being simple and fast to train compared to existing alternatives.