Multi-task Learning Based JPEG Steganalysis

Internet has become an important channel for human communication, and how to protect user data and communication security has been a problem in the field of information security, therefore steganography came into being.Steganography is the technique using insensitivity of the human perceptual system and redundancy of multimedia files to embed secret information into the multimedia files on purpose of secret communication. However, criminals also use it to harm national security and social stability.So steganalysis, the technique of determining whether multimedia files containing secret information has developed.JPEG images, due to its excellent quality, have become widely used in the Internet and a variety of image acquisition devices. Therefore, JPEG image steganography and steganalysis has also become a hot research field of information security.This paper first introduces several typical JPEG image steganography and steganalysis algorithms, and then froms the framework of universal JPEG image steganalysis. Secondly, discuss the diversity of JPEG image and its effect on steganalysis. Experiments show that the present pooled training steganalysis framework ignores the differnence of statistics characteristics to JPEG images of different sources and quantization tables.Then, according to the different performance of JPEG images of different sources and quantization tables, the paper presents the personalized training steganalysis framework of diverse JPEG images. Experiments show that when the training samples are sufficient, personalized training increases compared to pooled training, but when lacking of training samples, the personalized training can not distinguish stego images and cover images.Finally, due to the problem of personalized training, this paper proposes a multi-task learning based steganalysis framework by sharing the common different statistical characteristics of JPEG image of different sources and quantization tables. Experiments show that for JPEG images of same cameras and different quantization tables or of different cameras and different quantization tables, the detection results of multi-task learning based steganalysis are higher than pooled training and personalized training, and the improvement of the correct accuracy can be up to29.54%. When the accuracy of personalized training method is50.66%due to lacking of training samples, the multi-task learning method can achieve91.12%.