Personalized recommendation systems predict potential demand by analyzing user preferences. Generally, user feedback information is inferred from implicit feedback or explicit feedback. Nevertheless, feedback can be contaminated by user's mis-operations or malicious operations, and may thus lead to incorrect results. We propose a novel Multi-feedback pairwise ranking method via Adversarial training (AT-MPR) for recommender to enhance the robustness and overall performance in the event of rating pollution. The MPR method extends Bayesian personalized ranking (BPR) to cover three types of feedback: positive, negative, and unobserved. It obtains user preferences in a probabilistic way through multiple feedbacks at different levels. To reduce the impact of feedback noise, we train an MPR objective function using minimax adversarial training.