We investigate the use of complex network similarity for the identification of atrial fibrillation. The similarity of the network is estimated via the joint recurrence plot and Hamming distance. Firstly, we transform multi-electrodes epicardium signals recorded from dogs into the recurrence complex network. Then, we extract features representing its similarity. Finally, epicardium signals are classified utilizing the classification and regression tree with extracted features. The method is validated using 1000 samples including 500 atrial fibrillation cases and 500 normal sinus ones. The sensitivity, specificity and accuracy of the identification are 98.2%, 98.8% and 98.5% respectively. This experiment indicates that our approach may lay a foundation for the prediction of the onset of atrial fibrillation.