Stacked convolutional auto-encoders for steganalysis of digital images

In this paper, we point out that SRM (Spatial-domain Rich Model), the most successful steganalysis framework of digital images possesses a similar architecture to CNN (convolutional neural network). The reasonable expectation is that the steganalysis performance of a well-trained CNN should be comparable to or even better than that of the hand-coded SRM. However, a CNN without pre-training always get stuck at local plateaus or even diverge which result in rather poor solutions. In order to circumvent this obstacle, we use convolutional auto-encoder in the pre-training procedure. A stack of convolutional auto-encoders forms a CNN. The experimental results show that initializing a CNN with the mixture of the filters from a trained stack of convolutional auto-encoders and feature pooling layers, although still can not compete with SRM, yields superior performance compared to traditional CNN for the detection of HUGO generated stego images in BOSSBase image database.