eduzhai > Applied Sciences > Engineering >

BitMix Data Augmentation for Image Steganalysis

  • Save

... pages left unread,continue reading

Document pages: 3 pages

Abstract: Convolutional neural networks (CNN) for image steganalysis demonstrate betterperformances with employing concepts from high-level vision tasks. The majoremployed concept is to use data augmentation to avoid overfitting due tolimited data. To augment data without damaging the message embedding, onlyrotating multiples of 90 degrees or horizontally flipping are used insteganalysis, which generates eight fixed results from one sample. To overcomethis limitation, we propose BitMix, a data augmentation method for spatialimage steganalysis. BitMix mixes a cover and stego image pair by swapping therandom patch and generates an embedding adaptive label with the ratio of thenumber of pixels modified in the swapped patch to those in the cover-stegopair. We explore optimal hyperparameters, the ratio of applying BitMix in themini-batch, and the size of the bounding box for swapping patch. The resultsreveal that using BitMix improves the performance of spatial image steganalysisand better than other data augmentation methods.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...