Or bias the data in the training sets for malicious hidden behaviors.