Interpretable All-Type Audio Deepfake Detection with Audio LLMs via Frequency-Time Reinforcement Learning
Abstract
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.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
The main contribution of this paper is the introduction of a two-stage training paradigm for audio deepfake detection that enhances model interpretability and performance across diverse audio types. This work significantly advances the field by addressing the challenges of generalization and interpretability in deepfake detection, providing a robust framework that could be applied to future research and practical applications.
Methodology
The paper presents a novel approach to audio deepfake detection using audio large language models (ALLMs) through a two-stage training paradigm, FT-GRPO. The methodology is well-structured, beginning with the construction of a large dataset of rationales via an automatic annotation and polishing pipeline. This is followed by a cold-start supervised fine-tuning (SFT) phase and a reinforcement fine-tuning (RFT) phase that incorporates frequency-time constraints. The use of chain-of-thought (CoT) rationales is innovative, as it aims to enhance interpretability while addressing the limitations of traditional SFT approaches that reduce models to black-box classifiers. The integration of non-think samples to improve model robustness is also a notable contribution.
Results & evaluation
The experiments are comprehensive, evaluating the proposed method against multiple audio types (speech, sound, singing, music) and demonstrating state-of-the-art performance. The authors provide detailed accuracy results across different models and training types, showcasing the effectiveness of their approach. However, the paper could benefit from a more extensive discussion of the datasets used and the specific metrics employed for evaluation beyond accuracy, such as precision, recall, or F1-score.
Reproducibility
The paper includes a URL for accessing the code and data, which is crucial for reproducibility. The implementation details are described, including hyperparameters and training configurations. However, the paper could improve by providing a more detailed explanation of the experimental setup, such as the specific hardware used and any software dependencies.
Limitations
The authors acknowledge several limitations, including the focus on a limited set of ALLMs, potential suboptimal annotations, and the non-exhaustive nature of the testing conditions. These limitations may affect the generalizability of the findings and the robustness of the model in real-world scenarios.
Broader impact
The research addresses a critical issue in the realm of audio deepfake detection, which has significant implications for media integrity, cybersecurity, and misinformation. By improving the interpretability and accuracy of detection methods, this work could contribute to the development of more reliable systems for identifying synthetic audio, thereby enhancing trust in audio content across various applications. The main contribution of this paper is the introduction of a two-stage training paradigm for audio deepfake detection that enhances model interpretability and performance across diverse audio types. This work significantly advances the field by addressing the challenges of generalization and interpretability in deepfake detection, providing a robust framework that could be applied to future research and practical applications.
Author of this paper? Send a correction.