Reproduce “SilentCipher: Deep Audio Watermarking” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Mayank Kumar Singh, Naoya Takahashi, Weihsiang Liao, Yuki Mitsufuji

Sony AI's SilentCipher system introduces a deep learning method for audio watermarking that ensures embedded messages are imperceptible, robust against diverse distortions including non-differentiable compressions, and scalable to professional 44.1kHz sampling rates. It achieved high decoding accuracy, often 99-100%, across various attack scenarios and maintained an average Signal-to-Distortion Ratio of 47.24dB, with expert listeners unable to detect watermarks.

Key insights

  • Direct control over the Signal-to-Distortion Ratio (SDR) through a scaling factor simplifies the deep learning training process by decoupling perceptual quality from trainable parameters, ensuring a guaranteed quality floor without a separate perceptual loss.
  • Explicitly embedding psychoacoustic constraints, such as phase and magnitude relations, is essential for designing truly imperceptible deep learning watermarks, particularly for challenging audio segments like band-limited signals.
  • A pseudo-differentiable compression layer provides an effective generalizable technique to train deep learning models for robustness against real-world non-differentiable audio transformations.
  • The model's design, which concentrates message information into lower half-frequencies, contributes significantly to robustness against resampling attacks.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.