Distribution-Based Adversarial Multiple-Instance Learning