Almost every defence built against deepfakes so far arrives too late to matter. Detectors examine a file after it is posted, a clip after it circulates, a recording after the money has moved. The fraud that actually costs companies happens live, inside a call, while a person is talked into a transfer by a face they recognise. Researchers at the Fraunhofer Institute for Secure Information Technology in Darmstadt have built software that attacks the problem at that moment instead, scoring in real time how likely it is that the colleague on screen is synthetic.

Videoconferencing is a hostile environment for detection. The pixel-level irregularities that earlier forensic tools depended on are precisely what compression destroys, and network quality fluctuates enough that the same stream degrades differently minute to minute. The Fraunhofer team, working with the Fraunhofer Heilbronn Research and Innovation Center for Cybersecurity inside an ATHENE project, handled this by folding the distortion into training rather than filtering it out, simulating compression, quality variation and blur filters and generating their own deepfakes so the models would learn what a faked face looks like after a codec has been through it. Machine learning proved substantially faster and more efficient than classical signal-processing approaches. The audio side was trained on roughly 19,000 authentic recordings against about 160,000 manipulated ones from public datasets, and video and audio detection run together rather than separately.

The design constraints are as notable as the detection itself. Analysis runs locally and nothing leaves the machine, which keeps it inside data-protection requirements that would otherwise make scanning a business call untenable, and it needs only a computer with a modern graphics card and twelve gigabytes of graphics memory. What it produces is a running likelihood rather than a verdict. “If a conversation currently exhibits many indicators, the system outputs a probability of whether it could be a deepfake,” says Martin Steinebach, who heads Media Security and IT Forensics at Fraunhofer SIT. The team is explicit that the number is a prompt for a human step, not a replacement for one: ask a question only the real person could answer, or call back on a different channel.

The threat model behind this is no longer hypothetical. In the Arup case, an employee joined a video call populated by a synthetic CFO and synthetic colleagues and then moved about 25.6 million dollars across fifteen transactions to five Hong Kong accounts. Entrust, whose fraud report covers more than a billion identity verifications, found injection attacks up forty percent year over year, deepfaked selfies up fifty-eight percent through 2025, and a deepfake involved in roughly one in five biometric fraud attempts. Fraunhofer, citing Singapore’s cybersecurity agency, points to more than 1,200 video fraud attempts recorded there in January 2026 against 43 a year earlier.

What has not been established is how well the thing works. The demonstrator is a proof of concept, and no accuracy rate, false-positive rate or peer-reviewed paper has been published alongside it, leaving the important question open: a score that cries wolf during ordinary calls with bad lighting and worse bandwidth would be abandoned within a week. The next phase depends on companies and videoconferencing providers agreeing to integrate it, and on unresolved legal questions about consent, since analysing a call means analysing everyone in it. There is also the structural problem no detector escapes: it has learned the generators that exist now. Real-time warning is the right place to intervene. Whether a warning that arrives mid-sentence changes what a finance employee does next is a separate question, and nobody has tested it yet.


Discover more from TekCrispy

Subscribe to get the latest posts sent to your email.

Leave a comment

Leave a Reply