Audio Watermarking with Deep Neural Network and Perceptual Loss

Abstract

Digital watermarking consists in hiding a message in a digital media, image or audio, without affecting the user's experience. Watermarking can be used to deal with authorship rights and to guarantee a fair distribution of the royalties. Musical audio files can be subject to many modifications such as digital compression, slight change of pitch or speed, or noise addition, and the inserted watermark must remain after those transformations. In this thesis, the deep neural model, inspired by recent solutions, is composed of two convolutional networks: an auto-encoder inserts the watermark and a classifier retrieves the inserted binary message. The model is trained in a way that takes into account psychoacoustics and the masking effect of the human ear. A specific loss, based on the noise-to-mask ratio, has been designed for that goal. The power of the watermark noise is stronger than with previous solutions, but still, the watermark should not be perceived. The neural model also learns to retrieve the watermark from any location without any need for synchronization.