Improving Costs and Robustness of Machine Learning Classifiers Against Adversarial Attacks via Self Play of Repeated Bayesian Games