Open-Set Source Tracing of Audio Deepfake Systems

TL;DR

This paper addresses the challenge of open-set source tracing for audio deepfake systems, proposing a novel energy score adaptation and training methods that significantly improve detection performance against unseen deepfake sources.

Contribution

It introduces SME, a new energy score for out-of-distribution detection, and demonstrates its effectiveness in open-set audio deepfake source tracing.

Findings

SME improves FPR95 by 31% over traditional methods

SME-guided training reduces FPR95 to 8.3%

Augmentation techniques enhance open-set detection robustness

Abstract

Existing research on source tracing of audio deepfake systems has focused primarily on the closed-set scenario, while studies that evaluate open-set performance are limited to a small number of unseen systems. Due to the large number of emerging audio deepfake systems, robust open-set source tracing is critical. We leverage the protocol of the Interspeech 2025 special session on source tracing to evaluate methods for improving open-set source tracing performance. We introduce a novel adaptation to the energy score for out-of-distribution (OOD) detection, softmax energy (SME). We find that replacing the typical temperature-scaled energy score with SME provides a relative average improvement of 31% in the standard FPR95 (false positive rate at true positive rate of 95%) measure. We further explore SME-guided training as well as copy synthesis, codec, and reverberation augmentations, yielding an FPR95 of 8.3%.