XAttnMark: Learning Robust Audio Watermarking with Cross-Attention

TL;DR

XAttnMark tackles the challenge of robustly detecting and attributing watermarks in audio subjected to a wide range of edits, including generative modifications. It combines partial parameter sharing between the watermark generator and detector with a cross-attention decoding head that leverages a shared embedding table, plus a temporal conditioning module to diffuse the watermark over time. A psychoacoustic-aligned temporal-frequency masking loss guides imperceptible embedding, improving perceptual quality while maintaining high detection and attribution performance. The approach achieves state-of-the-art results across standard and generative edits and demonstrates strong generalization, enabling practical provenance tracking and copyright protection in AI-assisted audio content.

Abstract

The rapid proliferation of generative audio synthesis and editing technologies has raised significant concerns about copyright infringement, data provenance, and the spread of misinformation through deepfake audio. Watermarking offers a proactive solution by embedding imperceptible, identifiable, and traceable marks into audio content. While recent neural network-based watermarking methods like WavMark and AudioSeal have improved robustness and quality, they struggle to achieve both robust detection and accurate attribution simultaneously. This paper introduces Cross-Attention Robust Audio Watermark (XAttnMark), which bridges this gap by leveraging partial parameter sharing between the generator and the detector, a cross-attention mechanism for efficient message retrieval, and a temporal conditioning module for improved message distribution. Additionally, we propose a psychoacoustic-aligned temporal-frequency masking loss that captures fine-grained auditory masking effects, enhancing watermark imperceptibility. Our approach achieves state-of-the-art performance in both detection and attribution, demonstrating superior robustness against a wide range of audio transformations, including challenging generative editing with strong editing strength. The project webpage is available at https://liuyixin-louis.github.io/xattnmark/.

Figures (10)

• Figure 1: Quality-attribution performance trade-off curve across different watermarking strengths and the overall performance comparison on detection and attribution tasks. Higher values on both axes indicate better performance.
• Figure 2: System Diagram for XAttnMark.XAttnMark consists of a watermark generator and a watermark detector, with a shared embedding table that facilitates message decoding through a cross-attention module. In the generator part, we first employ an encoder network to encode the audio latent and then apply a temporal modulation to hide the message. The modulated latent is then fed into a decoder to produce the watermark residual. In the detector part, a linear detection head is used for detecting the presence of watermarks, and a cross-attention module with the shared embedding table is used for message decoding.
• Figure 3: Attribution accuracy with different #Users.
• Figure 4: Ablation study on the proposed architecture.
• Figure 5: MUSHRA subjective listening test results comparing perceptual quality across different watermarking methods. Higher scores indicate better audio quality as rated by human listeners. Our method achieves quality scores comparable to AudioSeal.