On the Traces of Audio Deepfakes: Analysis of Audio Components for Detection
| Author | Schäfer, Karla; Frick, Raphael |
|---|---|
| Date | 2026 |
| Type | Conference Paper |
| Abstract | Advances in audio deepfake generation have resulted in the ongoing challenge of detecting such audio, with the explainability of detection results being an area that has received comparatively little research attention. The present study investigates the efficacy of various segments of the audio signal as features for the detection of audio deepfakes. We examine multiple feature inputs, including foreground versus background, harmonic versus percussive, diverse frequency ranges, and room impulse response, to evaluate their effectiveness for deepfake detection. The training was conducted on the ASVspoof 2019 LA train set, and the evaluation was performed on the LA eval set for in-domain tasks, and on the in-the-wild and DiffSSD datasets for out-of-domain tasks. Using lower frequency ranges (0-4kHz) results in lower EERs compared to the full spectrogram. Furthermore, removal of room impulse response together with using MFCC as feature representation led to an enhanced EER on the in-the-wild test set (24%), compared to MFCC on the original audio (EER: 69.08%). |
| Conference | European Signal Processing Conference 2026 |
| Url | https://publica.fraunhofer.de/handle/publica/527334 |
Vulnerabilities and Security Gaps
Continuing Education
Find out about education and training opportunities offered by the institutions participating in ATHENE