Training for Faster Adversarial Robustness Verification via Inducing ReLU Stability

TL;DR

This paper introduces a training approach that enhances neural network robustness and significantly reduces verification complexity by promoting weight sparsity and ReLU stability, applicable across various training and verification methods.

Contribution

The paper proposes a novel co-design training method that improves both the robustness and verifiability of neural networks through targeted properties like weight sparsity and ReLU stability.

Findings

Improving weight sparsity makes verification problems more tractable.

Enhancing ReLU stability speeds up verification by 4-13 times.

Method is compatible with diverse training and verification techniques.

Abstract

We explore the concept of co-design in the context of neural network verification. Specifically, we aim to train deep neural networks that not only are robust to adversarial perturbations but also whose robustness can be verified more easily. To this end, we identify two properties of network models - weight sparsity and so-called ReLU stability - that turn out to significantly impact the complexity of the corresponding verification task. We demonstrate that improving weight sparsity alone already enables us to turn computationally intractable verification problems into tractable ones. Then, improving ReLU stability leads to an additional 4-13x speedup in verification times. An important feature of our methodology is its "universality," in the sense that it can be used with a broad range of training procedures and verification approaches.