This paper describes our submitted DKU-CMRI system for the ASVspoof2021 challenge. For the PA task, assuming that the vocoder can partially eliminate the replay channel information, we used the difference between the vocoder filtered audio and the original audio as the input feature and adopt multiple outlier detection models as the backend classifier. In addition, the pyroomacoustic toolkit and speed perturbation are applied to enhance the system performance. For the LA task, we adopt the Opus codec and the SoX toolkit to augment the training data, and RawNet2 and LFCC-LCNN models are utilized to determine the spoof/bona fide audio. In the evaluation phase, the proposed methods achieve 0.6824 min t-DCF and 24.25% EER on the PA task, and 0.3310 min t-DCF and 8.23% EER on the LA task, respectively. Experimental results demonstrate the strong robustness and generalization ability of our submitted system for the PA task.