ML Security PapersInterpretable All-Type Audio Deepfake Detection with Audio LLMs via Frequency-Time Reinforcement Learning
Yuankun Xie 1,2, Xiaoxuan Guo 1,2, Jiayi Zhou 2, Tao Wang 2, Jian Liu 2, Ruibo Fu 3, Xiaopeng Wang 3, Haonan Cheng 1, Long Ye 1
Published on arXiv
2601.02983
Output Integrity Attack
OWASP ML Top 10 — ML09
Key Finding
FT-GRPO achieves 99.75% accuracy on ASVspoof2019LA and 90.10% average accuracy across all audio types while producing interpretable frequency-time grounded rationales.
FT-GRPO (Frequency Time-Group Relative Policy Optimization)
Novel technique introduced
Recent advances in audio large language models (ALLMs) have made high-quality synthetic audio widely accessible, increasing the risk of malicious audio deepfakes across speech, environmental sounds, singing voice, and music. Real-world audio deepfake detection (ADD) therefore requires all-type detectors that generalize across heterogeneous audio and provide interpretable decisions. Given the strong multi-task generalization ability of ALLMs, we first investigate their performance on all-type ADD under both supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT). However, SFT using only binary real/fake labels tends to reduce the model to a black-box classifier, sacrificing interpretability. Meanwhile, vanilla RFT under sparse supervision is prone to reward hacking and can produce hallucinated, ungrounded rationales. To address this, we propose an automatic annotation and polishing pipeline that constructs Frequency-Time structured chain-of-thought (CoT) rationales, producing ~340K cold-start demonstrations. Building on CoT data, we propose Frequency Time-Group Relative Policy Optimization (FT-GRPO), a two-stage training paradigm that cold-starts ALLMs with SFT and then applies GRPO under rule-based frequency-time constraints. Experiments demonstrate that FT-GRPO achieves state-of-the-art performance on all-type ADD while producing interpretable, FT-grounded rationales. The data and code are available online.
Key Contributions
- Automatic annotation-and-polishing pipeline that constructs ~340K frequency-time structured chain-of-thought (CoT) rationales for audio deepfake detection datasets
- FT-GRPO: a two-stage training paradigm combining SFT cold start with GRPO under rule-based frequency-time domain constraints to improve interpretability and generalization
- State-of-the-art all-type audio deepfake detection achieving 99.75% accuracy on ASVspoof2019LA and 90.10% average accuracy across all audio types
🛡️ Threat Analysis
Primary contribution is a novel detection architecture for AI-generated audio content (deepfake speech, singing voice, environmental sounds, music) — squarely within AI-generated content detection under Output Integrity. The paper introduces FT-GRPO, a new training paradigm, not a domain application of existing methods.