OSIP: Tightened Bound Propagation for the Verification of ReLU Neural Networks

Abstract

Abstraction-based methods for the verification of ReLU-based neural networks suffer from rapid degradation in their effectiveness as the neural network’s depth increases. We propose OSIP, an abstraction method based on symbolic interval propagation in which the choice of the ReLU relaxation at each node is determined via optimisation. We present an implementation of OSIP on top of Venus, a publicly available toolkit for complete verification of neural networks. In the experiments reported, OSIP calculated bounds that were tighter than the state-of-the-art on ReLU networks from the first competition for neural network verification. As a case study we apply the method for the verification of VGG16, a deep, high-dimensional, 300,000 node-strong model used for object classification in autonomous vehicles against local robustness properties. We demonstrate that OSIP could verify the correctness of the model against perturbations that are larger than what can be analysed with the present state-of-the-art.