Detecting Audio Deepfakes on the Edge:Lightweight SSL-Based Detection in a Browser Plugin

View on arXiv ← Back to list

Authors: Octavian Pascu, Dan Oneata, Horia Cucu, Nicolas M. Muller

Published: 2026-06-29 18:09:27+00:00

AI Summary

This paper proposes an on-device audio deepfake detection model that offers enhanced privacy by performing local processing. The solution leverages a truncated self-supervised backbone with a simple logistic classifier, demonstrating improved accuracy and inference speed compared to existing cloud-based alternatives. The model is integrated into a browser plugin for user-friendly and secure deepfake detection.

Abstract

Audio deepfakes are a growing challenge for the general public, as well as for journalists and fact-checkers. The latter need reliable tools to verify the authenticity of their sources, while at the same time keeping their information private. Commercial deepfake detection solutions rely on cloud-based processing, which raises privacy concerns. To solve this problem, we propose an on-device audio deepfake detection model. We show that a truncated self-supervised backbone with a simple logistic classifier is both very fast and often more accurate than existing solutions. Our solution outperforms the baseline AASIST by 10% and improves inference speed by 40%. We integrate this model into a browser plug-in, which allows journalists and fact-checkers to detect deepfakes easily and securely. Code for the plugin is available at https://github.com/OctavianPascu97/Audio-Deepfakes-Browser-Plugin.

Key findings
The truncated Wav2Vec2-300m model at layer 7 achieved the lowest average Equal Error Rate (EER) of 8.4% across out-of-domain datasets, outperforming larger models and traditional CNN-based architectures. This approach improved inference speed by 40% and provided a viable compromise between accuracy and computational efficiency for real-time, on-device applications.
Approach
The authors propose using a truncated self-supervised learning (SSL) backbone, specifically the Wav2Vec2 XLS-R-300M model, with only its initial layers, combined with a linear classifier. This approach aims to achieve a balance between accuracy and computational efficiency, enabling on-device execution within a browser plugin for privacy-preserving deepfake detection.
Datasets
ASVspoof19, ASVspoof 2021 DF, Fake Or Real (FoR), MLAAD, In the Wild (ITW), TIMIT-TTS, WaveFake
Model(s)
Wav2Vec2 XLS-R-300M (truncated at layer 7)
Author countries
Romania, Germany
← Previous
Next →