Asymmetric Phase-Encoded Audio Watermarking

The researchers propose asymmetric phase encoding (APC), a training-free audio cryptographic signature layer for addressing the threat of deepfake audio to voice authentication systems, usable as a standalone provenance primitive or stacked with learned watermarks. The method combines Ed25519 digital signatures, Reed-Solomon error correction, pseudorandom STFT phase bin selection, and redundant QIM modulation on the amplitude difference of adjacent bin pairs, yielding a compact, non-repudiable, blind-extractable watermark. Tested on 1,000 LibriSpeech test-clean audio clips (each 10 seconds, 44.1 kHz) under eight attack configurations—covering end trimming, 8 kHz low-pass, 16 kHz resampling round-trip, and FLAC, MP3, and OGG-Vorbis re-encoding—the cryptographic verification rate under each condition is between 97.5% and 98.3%, with an average PESQ of 3.02 and CPU latency of tens of milliseconds.