An RNN-Based Framework for the MILP Problem in Robustness Verification of Neural Networks

Hao Xue, Xia Zeng, Wang Lin, Zhengfeng Yang, Chao Peng, Zhenbing Zeng; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 1842-1857

Robustness verification of deep neural networks is becoming increasingly crucial for their potential use in many safety-critical applications. Essentially, the problem of robustness verification can be encoded as a typical Mixed-Integer Linear Programming (MILP) problem, which can be solved via branch-and-bound strategies. However, these methods can only afford limited scalability and remain challenging for verifying large-scale neural networks. In this paper, we present a novel framework to speed up the solving of the MILP problems generated from the robustness verification of deep neural networks. It employs a semi-planet relaxation to abstract ReLU activation functions, via an RNN-based strategy for selecting the relaxed ReLU neurons to be tightened. We have developed a prototype tool L2T and conducted comparison experiments with state-of-the-art verifiers on a set of large-scale benchmarks. The experiments show that our framework is both efficient and scalable even when applied to verify the robustness of large-scale neural networks.