“State-of-the-art formal verification methods for neural networks are limited to verifying local adversarial robustness for small image classifiers on datasets like MNIST and CIFAR.”