Adversarial Learning Bounds for Linear Classes and Neural Nets