FRELA: Frequency-Layered Audio Watermarking for Robust Content Authentication

Linh Thao Nguyen, Duy Pham, Sijia Liu, Qiben Yan

Keywords: Audio watermarking, frequency-layered watermarking, voice cloning prevention, signal processing, audio security, copyright protection

TL;DR: We introduce Frela, a frequency-layered audio watermarking method that enhances robustness and audio quality by distributing the watermark across multiple spectral layers.

Abstract: In today’s world, the published audio content could be easily found at the risk of misuse, including voice cloning or generative model synthesis. This poses ethical and legal threats to individuals, creators, and organizations. Audio watermarking addresses these threats by embedding imperceptible identifiers in the audio. In this paper, we introduce FRELA, a frequency-layered audio watermarking method that distributes the localized watermark across multiple layers of the frequency domain. Each layer is watermarked on the basis of its spectral properties, enabling some to resist audio attacks while others may degrade. This layered redundancy enables partial watermark recovery even when the host signal is distorted. Experimental results demonstrate that FRELA preserves high audio quality while significantly enhancing robustness compared to existing techniques, even under adverse conditions including noise, speed and volume variations, dynamic changes, sampling distortions, temporal modifications, Encodec compression, and most notably, pitch shifting and frequency band filtering. Furthermore, FRELA remains effective in scenarios involving lossy transmission or partial signal degradation, making it suitable for real-world applications in copyright protection and content authentication.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 10073