Browser Plugin Detects Audio Deepfakes 40% Faster While Preserving Privacy
On-device model beats cloud-based solutions by 10% accuracy without exposing data.
A team of researchers (Octavian Pascu, Dan Oneata, Horia Cucu, Nicolas M. Muller) published a paper proposing a privacy-preserving audio deepfake detector that runs locally in a browser plugin. The solution uses a truncated self-supervised learning (SSL) backbone paired with a simple logistic classifier, achieving both high accuracy and low latency. It outperforms the current state-of-the-art AASIST model by 10% in detection accuracy while speeding up inference by 40%. This makes real-time verification feasible without sending sensitive audio to external servers.
The browser plugin is designed for journalists and fact-checkers who need to verify source authenticity while keeping their data private. By running on-device, the tool eliminates the privacy concerns inherent in cloud-based commercial deepfake detectors. The model is lightweight enough to run in a browser extension, offering a practical, secure alternative for everyday use. The research is published on arXiv (2606.30780) with code available on GitHub, enabling broader adoption and further development.
- Outperforms AASIST baseline by 10% in accuracy while reducing inference time by 40%.
- Runs entirely on-device in a browser plugin, eliminating cloud privacy risks.
- Uses a truncated self-supervised backbone with a logistic classifier for lightweight detection.
Why It Matters
Gives journalists a private, fast tool to verify audio authenticity without relying on cloud services.