Safety Verification of Deep Neural Networks ∗