Scalable Inference of Symbolic Adversarial Examples