Light Multiscale Conventional Neural Network for MP3 Steganalysis

In this paper, we propose a light multiscale convolution neural network to detect adaptive MP3 steganography, which can be used in attacking both the MP3 steganography based on Huffman codes substitution and the method through modifying sign bit in MP3 encoding. Especially, we decrease the model size and the occupation of graphics memory based on convolution factorization. At the same time, the convolution kernels with different size are applied in one layer, which is conducive to the retaining of the detection performance. And refer to the residual structure, a shortcut connection is used in the proposed network to enhance the performance of the network. The experimental result shows the accuracy can reach more than 90% when the payload rate is high. And the model size is reduced by 70% than the previous networks.