Robust Audio Watermarking Based on Invertible Neural Network

Audio watermarking technology leverages the characteristics of the human auditory system and the original audio carrier to imperceptibly embed watermark information into the audio. Traditional watermarking algorithms employ signal processing techniques and are limited by the experience of the model designer. In contrast, deep learning–based neural network audio watermarking algorithms offer greater adaptability and versatility, and their robustness can be enhanced through simulated attacks, marking an important direction for future development in audio watermarking technology. Related research primarily focuses on balancing the imperceptibility, robustness, and embedding capacity of watermark information. The audio watermarking model designed in this paper emphasizes imperceptibility and robustness. Imperceptibility is enhanced by designing a discriminator that ensures the human ear cannot distinguish between the original and watermarked audio. Robustness is improved by developing a simulated attack block, which provides strong resistance against multiple types of attacks, and by mitigating the damage caused by the attack layer through the invertible design of a neural network assisted by a balancing block. This study achieves high imperceptibility and strong robustness based on an invertible neural network. The experimental results demonstrate that the model performs well in terms of both watermark embedding and extraction accuracy, as well as anti-attack performance.