Convolutional neural networks (CNN) for image steganalysis demonstrate better performances with employing concepts from high-level vision tasks. The major employed concept is to use data augmentation to avoid overfitting due to limited data. To augment data without damaging the message embedding, only rotating multiples of 90 degrees or horizontally flipping are used in steganalysis, which generates eight fixed results from one sample. To overcome this limitation, we propose BitMix, a data augmentation method for spatial image steganalysis. BitMix mixes a cover and stego image pair by swapping the random patch and generates an embedding adaptive label with the ratio of the number of pixels modified in the swapped patch to those in the cover-stego pair. We explore optimal hyperparameters, the ratio of applying BitMix in the mini-batch, and the size of the bounding box for swapping patch. The results reveal that using BitMix improves the performance of spatial image steganalysis and better than other data augmentation methods.
View on arXiv