Digital Image Steganalysis Based On Deep Learning

The popularization of mobile communication and internet technology has brought great convenience to people's communication,in the mean while,the problem of communication privacy has attracted more and more attention.Therefore,as a technology aiming for concealed and secure communication,steganography has received more and more interest.And its adversarial detection technology,steganalysis,has also attracted more and more researchers' attention.In recent years,with the combination of deep learning and steganalysis,the performance of steganalysis is getting better and better,but current researches on deep steganalysis models are mainly concentrated in performance promoting in matching conditions.Facing to cover-source mismatch and model compression,this paper studies the following aspects:In spatial domain image steganalysis scenario,to address the problem that cover-source mismatch will lead to sharp performance decline of deep steganalysis model,this paper proposes a solution based on domain adaptation,which reduces the difference between the model's feature distribution in the training set and the testing set,in the meanwhile maintains model's ability to distinguish cover and stego images.And through two methods,the J-Net model based on JMMD distance and the DAS model based on adversarial training,this idea is realized,which effectively improve the detection performance in cover-source mismatch scenario.In addition,this paper also uses A-distance to measure the texture complexity between different data-sets,and proves its effectiveness through experiments.In face of massive JPEG images,steganalysis requires not only good detection performance,but also good detection efficiency.In JPEG domain steganalysis scenario,this paper studies the model compression of deep steganalysis model.Aiming at the problem that knowledge distillation method only constrains the classification prediction output and cannot constrain internal features,this paper proposes to use KL divergence to measure and constrain the feature inputs before fully connected layers.Compared with the knowledge distillation method that only constrains prediction output,the proposed method can improve the performance of student model without increasing the amount of model parameters and flops in prediction phase.Furthermore,this paper combines knowledge distillation and deep separable convolution to compress classic JPEG model SRNet(M-SRNet).Compared with its teacher network SRNet,the detection accuracy of M-SRNet is reduced by less than 5% under0.4bpnc embedding rate,its parameters and flops are only 1/20 and 1/15 of SRNet,its time consuming in prediction phase is also significantly declined.In addition,this paper also studies the similarity of the disturbance to image between steganography and adversarial attack.Based on their similarity,this paper proposes to use deep steganalysis model to detect adversarial samples.The baseline model in spatial domain,Base-CNN,and the baseline model in JPEG domain,SRNet,both show good detection performance in adversarial sample detection,as well as pretty generalization performance among different data sets and different gradient based adversarial attack algorithms.