Multi-user watermarking based on paralleled invertible neural networks

Jinghui Shi, Zhiguang Yang, Li Li, Jingyu Ye, Zhicheng Dong, Hanzhou Wu

Abstract: With the rapid development of artificial intelligence technology, deep-faked audios have become increasingly prevalent in cyberspace, creating an urgent need for effective detection and tracing mechanisms. Watermarking offers an active defense strategy for identifying deep-faked audios, and the integration of multi-user watermarking further enhances its protection and tracking capabilities. By leveraging multi-bit watermarking technology, multiple independent watermarks can be embedded within a single audio file, providing unique identifiers for different stakeholders (e.g., lyricists, composers, platforms, etc.). This enables granular content protection and precise source tracing across diverse scenarios. For instance, users can upload audios embedded with both personal watermarks (e.g., lyricist and composer) and platform watermarks, ensuring better monitoring of audio dissemination and usage. Existing methods for embedding multi-user watermarks often require training multiple networks, which can lead to interference during the training process. To address this challenge, we propose an audio watermarking technique based on Paralleled Invertible Neural Networks (PINN). Our approach enables the splitting of a single network into multiple sub-networks for multi-user watermark embedding while requiring only one network to be trained. Specifically, we design a parallel-structured invertible neural network that is highly flexible, allowing the network to be divided into multiple sub-networks for watermark rewriting or multi-user watermark embedding. These embedded watermarks remain independent of each other, ensuring robust and interference-free operations. Experiments demonstrate the framework’s robustness, achieving an average BER of 1.8% and SNR of 26.5 dB when embedding four 32-bit watermarks, and also exhibits strong resistance to common distortions, such as noise addition, MP3 compression, and re-sampling. Furthermore, partial rewriting experiments highlight the flexibility of our approach, showing that individual watermarks can be rewritten through small-scale sub-networks without affecting others. This capability makes our method highly practical for real-world applications, offering a scalable and efficient solution for multi-user watermarking in audio content protection and source tracing.

External IDs:dblp:journals/sivp/ShiYLYDW25