Учёные создали систему обнаружения дипфейков в видеоконференциях в реальном времени

Researchers have created a synthetic interlocutor warning system for video conferences in real time — a response to the fact that deepfake technologies have learned to imitate voice and appearance so convincingly that standard visual inspection no longer works.

Why Video Calls Became a Target for Fraudsters

Video conferencing has become a fundamental business tool: interviews are conducted through it, deals are signed, and the identities of clients and partners are verified. Trust in a video call was traditionally considered higher than in an email or voice call: it seemed difficult to pretend to be another person while looking into a camera. This logic no longer holds today.

Modern deepfake technologies are capable of overlaying another person's face onto a video stream in real time while simultaneously cloning their voice — with a synthesis latency that has become practically imperceptible. As a result, a call participant looks into the "eyes" of their interlocutor, hears a familiar voice, and is at the same time communicating with a fully synthetic image generated by a neural network.

• Deepfakes spoof face and voice simultaneously, directly during the call
• Synthesis quality has reached a level where visual analysis is no longer sufficient
• Video conferences are a direct target: negotiations, identity verification, financial decisions
• Most existing deepfake detectors work with recordings, not in real time

How the Warning System Works

The system analyzes the incoming video and audio stream directly during the call and displays a warning if it detects signs of synthesis. The key task is not to replace human judgment, but to provide a timely signal before the user makes a decision based on trust in their "interlocutor."

Deepfake detectors look for characteristic traces of generation: desynchronization of facial expressions and speech, artifacts at the edges of the face and hairline, anomalies in eye movement, unnatural blinking, non-standard lighting patterns. The task is complicated by the fact that generator neural networks are constantly improving — a detector must operate under conditions of an unceasing technological arms race.

"Increasingly, we have to ask ourselves: are real eyes looking at us from the screen," states a

TechXplore AI article on the threat of deepfakes in video conferencing.

The warning must appear in the call interface in real time — otherwise it loses its practical value and becomes a tool for retrospective investigation rather than live protection.

Why the Solution Emerged Now

By 2025–2026, the quality of deepfakes made a sharp leap: available commercial and open-source tools allow convincing fakes to be created without specialized technical knowledge. The barrier to entry has dropped so low that the technology has ended up in the hands of a wide range of malicious actors — from telephone scammers to participants in corporate espionage.

AI-based deepfake detectors have existed for several years, but most of them operate in post-hoc mode: they analyze already-recorded video. A real-time system is a fundamentally different challenge: the algorithm must deliver a verdict in milliseconds, while the conversation is still ongoing and while the user has not yet made a key decision.

What This Means

The emergence of practical deepfake warning tools directly within the video call interface is an important step for corporate security and online identification. Businesses that use video conferencing for client verification, financial negotiations, or personnel decisions will find that such a tool closes a real vulnerability — one that the human eye can no longer close in the era of generative AI.