Digital Audio Watermarking Method Based on Singular Spectrum Analysis with Automatic Parameter Estimation Using a Convolutional Neural Network

Abstract

This paper proposes an audio watermarking method based on the singular-spectrum analysis (SSA) incorporating with a convolutional neural network (CNN) for parameter estimation. A watermark is embedded into an audio signal by modifying some part of its singular spectrum according to an embedding rule. Such a modified part affects both the robustness of the scheme and sound quality of watermarked signals, and it should be determined appropriately in order to balance the robustness and sound quality. In our previous work, we used a method based on a differential evolution (DE) algorithm to estimate the suitable part. However, it is a time-consuming approach. Therefore, in this work, we replace it with a CNN approach. A dataset used to train the CNN is constructed based on the DE. Experimental results show that the computational time is considerably reduced by 96,923 times. The average bit-error rate is 0.07 when there is no attack, and the sound quality of watermarked signals satisfies three objective evaluation metrics. Also, the proposed scheme could blindly extract the watermark due to the time efficiency of the CNN-based method.