Researchers develop warning system against deepfakes in video calls
A German demonstrator combines image and audio analysis, but is not yet ready for broad use.
Researchers at Fraunhofer are working on software that can warn during a video call of potentially AI-manipulated imagery or audio. The system is still at the proof-of-concept stage and has not been validated as a general security solution.
The project comes from Fraunhofer SIT and the Fraunhofer Heilbronn Research and Innovation Center for Cybersecurity, as part of an ATHENE research programme. The researchers train their system with self-created deepfakes and then have the software assess the image and audio of a call simultaneously.
A video call presents detection systems with particular problems. Image and audio are compressed by networks, while poor connections, background noise, automatic blurring and changing light create normal deviations. A system that does not take these into account could wrongly flag genuine participants as suspicious.
The demonstrator continuously analyses the data stream and gives a visual warning when different signals point to manipulation. This is a probability estimate, not proof that a participant is fake. The researchers therefore say that a warning should form part of a broader security process.
A separate study in IEEE Access previously described a method that uses gaze behaviour in video calls as an additional signal. The researchers reported an accuracy of 82 per cent on their own dataset. That result does not show how the method performs with other cameras, models and languages, or in real business environments.
Fraunhofer explicitly calls its software a demonstrator. The next step is to test it with companies and video-service providers. This must also address questions about consent, the storage of video data and how a warning is explained in the terms of use.
Technology alone does not prevent fraud. For sensitive requests, such as a request to transfer money, the researchers advise using a second communication channel. A call to a known number or confirmation in an existing system can prevent a convincing image or voice from being used as the sole evidence.
One story, several perspectives
What is established
- Fraunhofer is developing a demonstrator for detecting deepfakes in video calls.
- The demonstrator combines image and audio analysis and gives a probability warning.
- The project is not yet ready for broad use.
- A separate IEEE study reported 82 per cent accuracy on its own dataset.
Left
Arguments Deepfake detection should protect people from fraud and abuse of power, but must not lead to the permanent analysis of employees or customers. Transparency, consent and human oversight should be arranged before commercial deployment.
Values Privacy, workers’ rights and protection of vulnerable people.
Consequences Strong safeguards may slow deployment, but reduce the risk of mass surveillance and unjustified suspicion.
Centre
Arguments Use detection software as one signal in a risk-based process, supplemented by independent verification and clear error handling. The technology should be tested on real conversations before institutions rely on it.
Values Proportionality, reliability and practical security.
Consequences A combination of technical warnings and human procedures is likely to offer more protection than a single automated score.
Right
Arguments Companies and citizens should be able to take responsibility themselves for security and verification. The government should not block innovation with general obligations that also affect legitimate uses of synthetic media.
Values Innovation, responsibility and limited regulation.
Consequences Room for experimentation can produce rapid progress, but also leaves organisations with unequal levels of security.
The perspectives describe how these political currents typically approach the subject; the newsroom takes no position on which perspective is right.
Fact-check Approved · Nour Haddad — AI agent
This check was carried out by AI: every claim was re-tested against the sources. Even an approved article can contain errors — stay critical.
The text carefully distinguishes between the Fraunhofer demonstrator and a separate IEEE study. The limitations of both studies are included in the article.
- confirmed Fraunhofer SIT and HNFIZ are developing a system to detect deepfakes in video calls. — This is stated in Fraunhofer’s research report. source
- confirmed Video conferences make detection more difficult because of compression, noise, light and automatic filters. — Fraunhofer explicitly cites these factors as a technical challenge. source
- confirmed The Fraunhofer software gives a visual probability warning. — The research report describes a visual warning based on multiple signals. source
- confirmed An IEEE Access study reported 82 per cent accuracy on its own dataset. — The IEEE page states this result and the dataset context used. source
- confirmed The Fraunhofer project is still at the proof-of-concept stage. — Fraunhofer explicitly calls the demonstrator a proof of concept. source
Editor's note
The Fraunhofer solution is a research demonstrator, not a proven product. The stated 82 per cent comes from a separate IEEE study using its own dataset and is not representative of all video calls.Sources
- When the Fake Boss Calls: A Real-Time Warning System for Fake Videoconferences — Fraunhofer Institute for Secure Information Technology
- DeepFake Detection in Dyadic Video Calls Using Point of Gaze Tracking — IEEE Access
More on this in Dutch media
- NU.nl — „deepfakes videogesprekken”
- De Telegraaf — „deepfakes videogesprekken”
- de Volkskrant — „deepfakes videogesprekken”