NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·3h ago
AI Research
arXiv CS.AI·May 25·8 min read
XAttnMark: Learning Robust Audio Watermarking with Cross-Attention
TL;DR
Generative audio tools enable copyright infringement and deepfakes, but existing watermarking methods struggle to simultaneously optimize robust detection and accurate attribution (identifying the watermark source). XAttnMark combines parameter sharing between encoder/decoder components, cross-attention mechanisms for message retrieval, and psychoacoustic-aligned masking loss to achieve state-of-the-art detection and attribution performance against audio transformations.
✦ Why It Matters
Engineers can implement XAttnMark to protect audio content from unauthorized use and verify authenticity in generative AI applications.
Key Takeaways
- Use cross-attention for efficient watermark message retrieval in audio systems.
- Apply psychoacoustic masking loss to improve watermark imperceptibility to human listeners.
- Share parameters between encoder and decoder to jointly optimize detection and attribution.
- Test watermark robustness against generative editing attacks, not just standard audio transformations.
Full Summary
Audio watermarking embeds imperceptible identifying signals into sound files to protect intellectual property and verify authenticity. Prior neural methods like WavMark and AudioSeal improved robustness but faced trade-offs between reliable detection and accurate attribution (determining who created the watermark).
XAttnMark addresses this by using partial parameter sharing between the watermark generator and detector networks, enabling more efficient learning. A cross-attention mechanism—a neural component that selectively focuses on relevant information—retrieves embedded messages more accurately.
A temporal conditioning module distributes the watermark signal better across time. Critically, the authors introduced psychoacoustic-aligned time-frequency masking loss, which accounts for how human hearing perceives sound, making watermarks more imperceptible.
Testing showed superior robustness against various audio transformations and generative editing attacks.