A Mixed Integer Programming Approach for Verifying Properties of Binarized Neural Networks
This work addresses the verification bottleneck for neural networks, particularly in safety-critical domains like aircraft control, but it is incremental as it builds on existing BNN and verification methods.
The paper tackles the scalability problem of verifying neural networks by proposing a mixed integer programming formulation for binarized neural networks (BNNs), demonstrating runtime reductions compared to state-of-the-art verification algorithms for full-precision networks on MNIST and an aircraft collision avoidance controller.
Many approaches for verifying input-output properties of neural networks have been proposed recently. However, existing algorithms do not scale well to large networks. Recent work in the field of model compression studied binarized neural networks (BNNs), whose parameters and activations are binary. BNNs tend to exhibit a slight decrease in performance compared to their full-precision counterparts, but they can be easier to verify. This paper proposes a simple mixed integer programming formulation for BNN verification that leverages network structure. We demonstrate our approach by verifying properties of BNNs trained on the MNIST dataset and an aircraft collision avoidance controller. We compare the runtime of our approach against state-of-the-art verification algorithms for full-precision neural networks. The results suggest that the difficulty of training BNNs might be worth the reduction in runtime achieved by our verification algorithm.