Unsupervised Data Drift Monitoring for Deployed Deepfake Speech Detectors

Authors: Lukáš Prokeš, Vojtěch Staněk, Anton Firc, Kamil Malinka

Published: 2026-10-12 14:08:25+00:00

Comment: Accepted at IEEE International Workshop on Information Forensics and Security (WIFS) 2026

AI Summary

This paper proposes an unsupervised data drift monitor for deployed deepfake speech detectors to address the challenge of evolving attacks and conditions where labels are unavailable. The monitor uses regularized Mahalanobis distances on internal detector embeddings to categorize inputs into trusted, ambiguous, and novel regions. This approach substantially reduces the security-critical case of accepting deepfakes by routing unfamiliar inputs for review, highlighting that detector reliability is as important as binary classification accuracy.

Abstract

Deepfake speech detectors are typically evaluated as binary classifiers on fixed benchmarks, but deployed systems face evolving attacks, codecs, speakers, and recording conditions, with labels unavailable during operation. This makes it difficult to know when a detector operates in familiar conditions and can be relied on. We propose an unsupervised drift monitor that operates on the internal embeddings of a deployed detector. The monitor compares each input against bona fide and spoof references using regularized Mahalanobis distances. These distances define trusted, ambiguous, and novel regions of the detector embedding space, allowing unfamiliar inputs to be routed for review. We evaluate the monitor on four modern detectors on three datasets. The monitor raises alerts for dataset mismatches and substantially reduces the security-critical case of accepting deepfakes at the cost of reduced automatic coverage. The results also show that drift affects both classes: bona fide recordings can move away from the training reference, increasing alert rates even in in-domain settings. Our findings suggest that deployed deepfake speech detectors should be treated not only as a binary classification, but also as a reliability problem.

Key findings

The monitor effectively detects dataset mismatches and substantially reduces the false acceptance rate of deepfakes from critical levels (e.g., over 60%) to below 1% across various settings. It also reveals that data drift affects both spoof and bona fide recordings, leading to increased alert rates even for in-domain data due to evolving conditions like codecs and channels. Covariance regularization is crucial for stable Mahalanobis monitoring due to the anisotropic nature of detector embeddings.

Approach

The proposed monitor operates on the internal utterance embeddings of a fixed deepfake speech detector. It compares each input embedding against a bona fide and a spoof reference profile, constructed from training data using regularized Mahalanobis distances. These distances define trusted, ambiguous, and novel regions in the embedding space, allowing the system to alert and route inputs outside familiar conditions for review rather than blindly trusting the detector's binary output.

Datasets

ASVspoof 2019 LA (AS19), ASVspoof 5 (AS5), In-the-Wild (ITW)

Model(s)

AASIST, MHFA, XLS-R (as frontend)

Author countries

Czech Republic

← Previous