Improving Audio Steganalysis Using Deep Residual Networks

In this paper, we propose an effective audio steganalysis scheme based on deep residual convolutional networks in the temporal domain. Firstly, considering the weak difference between cover and stego, a high pass filter is adopted in the proposed network which is used to calculate the residual map of the audio signal. Then, comparing with convolutional neural networks (CNNs) based audio steganalysis in recent studies, the deeper network structure and complicated convolutional modules are considered to capture the complex statistical characteristic of steganography. Finally, batch normalization layers and shortcut connections are applied to decrease the dangers of over-fitting and accelerate the convergence of back-propagation. In the experiments, we compared the proposed scheme with CNNs based and hand-crafted features based audio steganalysis methods to detect the various steganographic algorithms on speech and music audio clips respectively. The experimental results demonstrate that the proposed scheme is able to detect multiple state-of-the-art audio steganographic schemes with different payloads effectively and outperforms several recently proposed audio steganalysis methods.