Vstorm’s engineer supports audio deepfake analysis – CVPR 2026

Deepfakes pose a growing threat to companies, individuals, and society at large. According to statistics compiled by Eftsure, as many as 60% of consumers encountered a deepfake video in the past year. The danger is real: deepfake-related scams cost companies nearly $450,000 on average.

The key challenge lies in telling authentic content apart from its deepfaked counterparts. Only 0.1% of people can correctly identify every deepfake they are shown, and the barrier to creating one is low. According to ElevenLabs, just 30 seconds of clean audio is enough to produce a high-quality voice clone. For public figures such as politicians or high-profile CEOs, voice samples are abundant.

Conducted research focuses on automating audio deepfake detection. A relatively unexplored area is source tracing: identifying the specific model used to generate a given piece of content. This information can be critical from a forensic point of view and it may also help harden models against misuse and make malicious actors easier to identify.

The paper was prepared by Vstorm engineer Dawid Wolkiewicz alongside Piotr Syga, both academically affiliated with the Wrocław University of Science and Technology.

A Hierarchical Margin-Based Approach for Audio Deepfake Source Tracing

To tackle source tracing, the researchers proposed “From Supra to Sub: A Hierarchical Margin-Based Approach for Audio Deepfake Source Tracing,” presented at the CVPR 2026 APAI Workshop.

Their key idea is that a synthetic voice leaves fingerprints on two levels. The broad one points to the family of systems it came from, think of well-known text-to-speech engines like Tacotron, VITS, or XTTS. The finer one points to the specific version within that family. The system works the same way a detective might: it first narrows down the general type of tool used, then zeroes in on the exact one.

What sets this approach apart is that it is built for messy real-world conditions. In practice, most deepfakes an investigator runs into were made with tools the system has never seen before. Rather than forcing a wrong guess, it can say “I don’t recognize this”, or give a partial answer, naming the broad family even when the exact version remains unknown. That honesty about uncertainty is exactly what a forensic tool needs.

Tested against a standard benchmark, the method identified sources noticeably more accurately than the leading existing approach, with the biggest gains on precisely those never-before-seen tools. It is also practical to maintain: when a new deepfake tool appears within a known family, only a small part of the system needs updating rather than the whole thing, which is an important advantage in a field where new generators emerge constantly.

Summary

This approach can help analysts map out coordinated deepfake campaigns and trace the malicious actors and institutions behind them.

The research was presented at the Authenticity & Provenance in the age of Generative AI (APAI) Workshop at the 2026 Conference on Computer Vision and Pattern Recognition (CVPR) in Denver, Colorado.