Recent studies have proved that CNN (Convolution neural network) is an effective method to solve steganalysis problem. In this paper, we analyze the influence of the input data sequence on CNN training results according to the training mechanism of CNN. The differences made by variant data sequence may be caused by data correlation, data mini-batch processing and high-pass filtering operations. In experiment, the accuracy of CNN detection with variant data sequence was compared, with cover and stego produced by two steganography algorithms (S_UNIWARD, WOW) with different embedding rates (0.1, 0.4). Results demonstrate that the network detection effect of pairwise data is better than that of random one.