SE-ADD: Self-Evolving Audio Deepfake Detection via Reward-Guided Forensic Cue Selection

The rapid evolution of deepfake audio increasingly challenges the existing detection methods, driving researchers to focus on adapting audio language models (ALMs) for audio deepfake detection (ADD). While previous studies treat ALMs as binary classifiers, we propose a self-evolving audio deepfake detection (SE-ADD) framework, which makes use of ALMs’ potential acoustic forensic reasoning capability through reward-guided self-generated forensic cue selection. For each audio training sample, an ALM is prompt to generate multiple candidate acoustic forensic cues, evaluated based on the resulting detection accuracy. High-utility audio-cue pairs are then aggregated to finetune the ALM, forming an iterative generation-selection-update cycle. Experiments on self-evolving Qwen2Audio across spoofed speech and general audio show that self-generated forensic cues substantially improve ALM-based detection, while successive evolution cycles in SE-ADD provide further gains.