Can External Sources Help the Knowledge Cut-off Issues in Audio Deepfake Detection?

Abstract

In conventional deepfake audio detection pipelines, inference and evaluation are typically conducted using standalone, parameterized models trained exclusively on a fixed dataset. In contrast, we investigate a retrieval-augmented detection framework that queries an external knowledge database at inference time-analogous to retrieval augmented generation in large language models-to mitigate knowledge cutoffs and to find the conditions under which this retrieval-augmented detector outperforms purely parameterized baselines. The external knowledge database can be updated continuously, including during system operation, but its optimal use scenarios have not been previously well characterized. Our empirical results indicate that retrieval augmentation is most beneficial when the content of the knowledge database is distinct from the fine-tuning data of self-supervised learning (SSL)-based models yet remains aligned with the domain of the evaluation set; under these circumstances, linear score-level fusion with standalone detectors is also effective. Nonetheless, the retrieval-augmented system remains vulnerable to domain shifts, and the performance improvements are limited when the test distribution diverges substantially from the domain represented in the knowledge database.

accepted by WIFS 2026