Robust and Verifiable Information Embedding Attacks to Deep Neural Networks via Error-Correcting Codes.