Robust watermarks for audio diffusion models by quadrature amplitude modulation

Abstract

Generative models enable the creation of high-quality digital content, including images, videos, and audio, making these tools increasingly accessible to users. As their use grows, so does the need for robust copyright protection mechanisms. Existing watermarking methods, primarily post-hoc, can safeguard the copyrights of users but fail to protect service providers, leaving room for intentional misuse, such as erasing watermarks and falsely claiming originality. To address this, previous works proposed integrating watermarks into the noise initialization of diffusion models for image generation, ensuring robustness against attacks like cut-and-paste. However, this approach has not been investigated for audio generation, where the 1D nature of audio data requires fundamentally different techniques. In this paper, we propose a novel barcode-like watermarking method for audio diffusion models, leveraging 4-quadrature amplitude modulation (4-QAM) to embed twice as much information as amplitude modulation methods for existing image generations. Our approach demonstrates significantly improved robustness against attacks, including cut-and-paste, and outperforms state-of-the-art audio watermarking techniques in preserving information and ensuring copyright protection for both users and service providers.