Incremental Verification of Neural Networks

Complete verification of deep neural networks (DNNs) can exactly determine whether the DNN satisfies a desired trustworthy property (e.g., robustness, fairness) on an infinite set inputs or not. Despite tremendous progress to improve scalability complete verifiers over years individual DNNs, they are inherently inefficient when deployed is updated its inference speed accuracy. The inefficiency because expensive verifier needs be run from scratch DNN. To efficiency, we propose new, general framework for incremental and based design novel theory, data structure, algorithms. Our contributions implemented in tool named IVAN yield overall geometric mean speedup 2.4x verifying challenging MNIST CIFAR10 classifiers 3.8x ACAS-XU state-of-the-art baselines.