Research On Steganalysis Method With Scalability And Few-sample Learning Abilit

Steganography is a technique to hide secret information unknowingly in multimedia carriers.As an adversary of steganography,steganalysis aims to determine the existence of secret information hidden in a multimedia carrier.To detect existing steganographic algorithms,an elaborate steganalysis model is usually trained on a dataset consisting of pairs of cover/stegoimages generated by these steganographic algorithms.However,when a new steganographic algorithm emerges,it is often necessary for the existing steganalysis model to completely retrain the model on both the previous steganographic dataset and the new steganographic algorithm dataset to obtain the detection capability of the old and new algorithms.With the continuous emergence of new steganographic algorithms,the above steganalysis models are inefficient and unrealistic,which makes them difficult to be applied in real-world scenarios.On the other hand,when new steganographic algorithms emerge,there is often a lack of sufficient data available for training against this algorithm,and the success of common deep neural network-based steganalysis models often relies on a large amount of data for training,and for the case of small sample datasets,the performance of the trained steganalyzer is not satisfactory therefore.Based on the above two shortcomings encountered by existing steganalysis models in the face of newly emerged steganographic algorithms,this paper proposes a steganalysis method with scalability and few-sample learning capability,and the detailed research results are as follows.(1)A steganalysis model based on continue learning is proposed.The model is first trained on the existing steganographic dataset,and when a new steganographic algorithm appears,this steganalysis model can be trained incrementally on the image dataset generated by the new steganographic algorithm.In this scheme,when a steganalysis model is trained on the new image dataset generated by the new steganographic algorithm,its network parameters are effectively and efficiently updated with sufficient consideration of their importance evaluated in the previous training process.This approach can guide the steganalysis model to learn the patterns of the new steganographic algorithm without significantly degrading the detectability against the previous steganographic algorithms.(2)A steganography analysis model based on contrastive learning is proposed.The model first generates multiple sets of additional available training samples by data augmentation when facing a steganographic dataset with few samples,and then draws the combined training data sample pairs from the extended dataset.By combining three loss functions to optimize the network,the network learns rich steganographic features quickly.This method reduces the reliance on labeled training data and achieves results comparable to traditional methods on fewer sample cases and fewer training rounds.