VoIP Steganalysis Using Shallow Multiscale Convolution and Transformer

Steganography is an effective method for transmitting secret information, but it can also be used for illegal activities such as terrorism, organized crime and data theft, etc. To solve the problem of steganography being used for malicious purposes, steganalysis technology has been developed. Steganalysis aims to detect whether the data has been steganography and identify whether it contains secret information, which is a kind of reverse process of steganography. VoIP data stream usually has high redundancy, which makes it an ideal carrier for steganography. In this paper, a Steganalysis Transformer (SAT) VoIP voice steganalysis method based on Transformer neural network is proposed with VoIP voice as the research object. The method first encodes the relative position of the features extracted from VoIP voice signals, combines the multi-scale convolution method to improve the local feature extraction to obtain more detailed feature information, transforms the high-dimensional sparse matrix into the low-dimensional dense features by mapping, and then realizes the steganalysis analysis through the feature extraction by the improved Transformer; the proposed SAT method is able to obtain the global features from the shallow layer and learn the high quality intermediate features. Experiments show that the SAT method proposed in this paper has superior performance, and the accuracy of VoIP steganalysis reaches 96.41%.