GenTraceBench: A Benchmark for Tracing Audio Deepfakes Across Pre- and Post-training Stages

View on arXiv ← Back to list

Authors: Li Wang, Kunyu Feng, Wan Lin, Dekun Chen, Qinke Ni, Xueyao Zhang, Lei Wang, Jie Shi, Haizhou Li, Zhizheng Wu

Published: 2026-09-18 13:15:52+00:00

Comment: 5 pages, 2 figures, 4 tables. Accepted to the 15th International Symposium on Chinese Spoken Language Processing (ISCSLP 2026)

AI Summary

This paper introduces GenTraceBench, a benchmark designed to evaluate the persistence of forensic fingerprints in audio deepfakes as Text-to-Speech (TTS) models undergo adaptation processes like fine-tuning and preference optimization. The benchmark covers five TTS architectures with 16 pre-/post-training variants, totaling 49,728 utterances, and uses a train-on-foundation, test-on-adapted protocol to assess binary detection, closed-set attribution, and open-set verification. The findings indicate that while some adaptation methods preserve fingerprints, others, particularly certain supervised fine-tuning (SFT) and pre-training data changes, lead to substantial drift, impacting the efficacy of deepfake detection systems.

Abstract

Modern text-to-speech (TTS) systems are rarely deployed as unchanged pre-trained models. They are often adapted through supervised fine-tuning (SFT) or preference optimization such as DPO and GRPO. This raises a practical question for audio deepfake forensics: do fingerprints learned from a foundation generator remain valid after adaptation? We present GenTraceBench, a controlled benchmark spanning five TTS architectures, 16 pre-/post-training variants, and 49,728 utterances generated with fixed texts and speaker prompts. Under a train-on-foundation, test-on-adapted protocol, we evaluate binary detection, closed-set attribution, and open-set verification. DPO and GRPO generally preserve fingerprints, whereas some SFT and pre-training-data changes cause substantial drift; effect sizes vary across three forensic backbones. Repeated training runs confirm the largest W2V-BERT attribution drop, while a data-mixture control with comparable speech quality shows that composition change need not cause drift. In W2V-BERT verification, multi-shot enrollment reduces EER for the SFT condition from 44.4% to 11.0%, whereas the SingNet-only condition remains at or above 45% EER.

Key findings

DPO and GRPO (preference optimization methods) generally preserve forensic fingerprints in generated audio, whereas certain SFT (supervised fine-tuning) and changes in pre-training data composition can cause significant fingerprint drift. Multi-shot enrollment can effectively reduce verification EER for some SFT conditions by averaging out within-class variability, but it is insufficient for conditions where there's a systematic shift in the mean representation, highlighting the need for updated forensic systems in such cases.

Approach

The authors developed a controlled benchmark, GenTraceBench, by generating speech from five TTS architectures under 16 different pre- and post-training conditions. They then evaluated forensic backbones (W2V-BERT, WavLM, AASIST) on three tasks: binary detection, closed-set attribution, and open-set verification, using a 'train-on-foundation, test-on-adapted' protocol to measure fingerprint stability. They analyze the impact of different adaptation strategies and pre-training data composition on the detectability and traceability of generated audio.

Datasets

Seed-TTS (for prompts), Emilia, SingNet, INTP, NVSpeech

Model(s)

Wav2Vec2-BERT (W2V-BERT), WavLM, AASIST

Author countries

China

← Previous