On Training and Verifying Robust Autoencoders