Deepfake-Resilient Semi-Fragile Audio Embedding with Adaptive Bandlimited Spread-Spectrum Embedding and Neural Provenance Verification

Abstract

The rapid proliferation of deepfake audio poses significant threats to digital security and media integrity. This paper proposes a resilient semi-fragile audio watermarking framework for speech provenance authentication. The proposed Adaptive Bandlimited Spread-Spectrum (ABSS) embedder is guided by a psychoacoustic masking function and paired with a CNN-Transformer Neural Provenance Verifier (NPV) conditions on Log-Mel spectrograms, CQCC features, and a watermark residual channel. Evaluation on three benchmarks ASVspoof 2021, ADD 2023, and WaveFake yields detection accuracies of 99.89%, 96.05%, and 100% respectively, with SNR exceeding 56 dB and PESQ above 4.59 on all datasets. Bit Error Rate (BER) values clustered near 0.5 under seven benign attack categories, showing that the ABSS watermark does not survive routine signal processing such as MP3 compression or resampling. We report this candidly as a robustness limitation of the current embedding configuration rather than evidence of semi-fragile behavior and clarify that the watermark residual channel functions as a complementary auxiliary input to the NPV rather than a standalone provenance signal. With Wilson-score 95% confidence intervals accompanying every reported accuracy, the CNN-Transformer NPV nonetheless achieves strong in-domain detection performance across all three benchmarks (99.89%, 96.05%, and 100%, respectively), though we caution that these results are obtained on non-official, comparatively small Kaggle-sourced test partitions rather than the official ASVspoof evaluation protocol.