A Mathematical Model for Optimum Error-Reject Trade-Off for Learning of Secure Classification Models in the Presence of Label Noise During Training.