Deep learning frameworks have recently achieved superior performance in manypattern recognition problems. However, adoption of deep learning in imagesteganalysis is still in its initial stage. In this paper we propose a hybriddeep-learning framework for JPEG steganalysis incorporating the domainknowledge behind rich steganalytic models. We prove that the convolution phaseand the quantization & truncation phase of the rich models are not learnable indeep convolutional neural networks. Based on theoretical analysis, our proposedframework involves two main stages. The first stage is hand-crafted,corresponding to the convolution phase and the quantization & truncation phaseof the rich models. The second stage is a compound deep neural networkcontaining three deep subnets in which the model parameters are learned in thetraining procedure. By doing so, we ably combine some merits of rich modelsinto our proposed deep-learning framework. We have conducted extensiveexperiments on a large-scale dataset extracted from ImageNet. The primarydataset used in our experiments contains 500,000 cover images, while ourlargest dataset contains five million cover images. Our experiments show thatthe proposed framework outperforms all other state-of-the-art steganalyticmodels either hand-crafted or learned using deep networks in the literature.Furthermore, we demonstrate that our framework is insensitive to JPEG blockingartifact alterations and the learned model can be easily transferred to adifferent attacking target. Both of these properties are of critical importancein practical applications. According to our best knowledge, This is the firstreport of deep learning in image steganalysis validated with large-scale testdata.
展开▼
