Hiding speech in music files

Abstract

In large-capacity audio steganography, how to reduce distortion of the steganographic audio and reconstruct the high-quality secret audio are two crucial issues. In this paper, we propose a new invertible audio steganography network, InvASNet, to conceal secret speech in music files. Firstly, we adopt an orthogonal module to decompose the audio into uncorrelated components. In such a way, we can constrain the embedding of the secret audio into the less perceptible high-frequency subband of the host audio, thereby minimizing potential distortion in the low-frequency subband. Secondly, we consider the concealment and recovery processes as a pair of reversible operations, and then introduce the forward and inverse processes of the invertible neural networks (INNs) to model them, respectively. Compared with existing methods based on convolutional neural networks, our approach possesses a highly reversible structure and can leverage the lost information effectively. Furthermore, to enhance the capability of reversible audio, we develop a feature fitting module to learn more adaptive weights and biases of mappings in INNs. Extensive experimental results show that the proposed InvASNet achieves superior imperceptibility and competitive security in large-capacity steganography.