ML Security PapersDisentangling Speaker Traits for Deepfake Source Verification via Chebyshev Polynomial and Riemannian Metric Learning
Xi Xuan 1,2, Wenxin Zhang 3, Zhiyu Li 4, Jennifer Williams 5, Ville Hautamäki 1, Tomi H. Kinnunen 1
1 University of Eastern Finland
2 City University of Hong Kong
3 University of Chinese Academy of Sciences
Published on arXiv
2603.21875
Output Integrity Attack
OWASP ML Top 10 — ML09
Key Finding
Demonstrates that speaker traits significantly entangle with source embeddings (speaker verification model achieves 29.42% EER on source verification task), and SDML framework effectively disentangles them
SDML (Speaker-Disentangled Metric Learning)
Novel technique introduced
Speech deepfake source verification systems aims to determine whether two synthetic speech utterances originate from the same source generator, often assuming that the resulting source embeddings are independent of speaker traits. However, this assumption remains unverified. In this paper, we first investigate the impact of speaker factors on source verification. We propose a speaker-disentangled metric learning (SDML) framework incorporating two novel loss functions. The first leverages Chebyshev polynomial to mitigate gradient instability during disentanglement optimization. The second projects source and speaker embeddings into hyperbolic space, leveraging Riemannian metric distances to reduce speaker information and learn more discriminative source features. Experimental results on MLAAD benchmark, evaluated under four newly proposed protocols designed for source-speaker disentanglement scenarios, demonstrate the effectiveness of SDML framework. The code, evaluation protocols and demo website are available at https://github.com/xxuan-acoustics/RiemannSD-Net.
Key Contributions
- Speaker-disentangled metric learning (SDML) framework using Chebyshev polynomial-based loss to stabilize disentanglement optimization
- Hyperbolic space projection with Riemannian metric distances (HAM-Softmax) to separate speaker and source embeddings
- Four new evaluation protocols for source-speaker disentanglement scenarios on MLAAD benchmark
🛡️ Threat Analysis
The paper addresses verifying the source (which generator) of synthetic speech outputs. Source verification is fundamentally about authenticating AI-generated content provenance — determining which model/generator produced a given synthetic audio sample. This is output integrity verification, not detecting whether audio is synthetic (which would also be ML09 but a different flavor). The paper builds a system to trace deepfake audio back to its source generator, which is content provenance tracking.