ThiNet Based Pruning Method for GAN Based Steganography Framework UT-GAN

With the application of deep-learning framework, both the steganography and steganalysis gain superior performance than before, which has been testified by the previous researchers through experiments. Though the deep-learning models bring the excellent performance, they also result in the over-parameterized models and redundancy in computation and memory storage which make it hard for them to deploy. In this paper, we propose a ThiNet based scheme to prune the state-of the-art steganographic model based on GAN--UTGAN. We evaluate and sort the significance of the channels of a layer according to the output of the next layer by using the greedy algorithm and obtain the channels which have the similar impact on the next layer compared with all the channels before pruning. The experimental results conducted on SZUBase, BOSSBase, and BOWS2 dataset demonstrate that our proposed ThiNet based method could compact the original UT-GAN model effectively with about only two percent of parameters of the original model and have about only a little decline in detection error rate when analyzed by SRM with ensemble classifier, which has outperformed the ASDL-GAN method.