Early Warning System AI Detects Deepfakes in Video Conferences

Related Vendors

Fraunhofer-Institut für sichere Informations- technologie SIT

Taiwan Machine Tool & Accessory Builders' Association (TMBA)

SIMTOS (Seoul International Machine Tool Show)

Deepfakes do not stop at video conferences: voices and images can now be faked with astonishing realism. Fraunhofer researchers are therefore developing an AI-based system that detects video and audio deepfakes during a conversation and issues a warning if suspicion arises.

Deepfakes are increasingly being used for fraud and identity theft—including in video conferences. With the help of artificial intelligence, voices and faces can now be faked in real-time so realistically that even familiar conversation partners hardly suspect anything. In 2025, for instance, a finance director in Singapore participated in a video call where all participants—including his supposed superior—were AI-generated. He subsequently transferred nearly 500,000 USD. The amount was later recovered.The problem is increasing worldwide. In Singapore alone, more than 1,200 cases of video fraud were recorded in January 2026—compared to 43 cases in the same month the previous year. The Entrust Identity Fraud Report 2026 also warns that identity fraud is now industrialized, globally organized, and commercially optimized.

AI security

Five Protective Measures Against Social Engineering

Detect Manipulations Based on Image and Audio Data

Fraunhofer researchers are therefore working on a solution to detect deepfakes in video conferences as early as possible. A team led by Prof. Martin Steinebach from the Fraunhofer Institute for Secure Information Technology SIT combines the analysis of visual and audio data for this purpose. The AI-based software continuously analyzes the conversation and provides a probability rating when multiple signs of manipulation are detected.
The detection is particularly challenging because video conferences often have fluctuating image and sound quality. Compression, changing lighting, background noise, movements, and automatic blurring can cause image errors that must not be mistaken for a deepfake. For this reason, the researchers are training the system with real and manipulated recordings that are additionally equipped with typical disturbances from video conferences.

Deepfake Detection Directly on the Device

The analysis is designed to run locally on a powerful laptop. This eliminates the need to transmit image and audio data to an external server. This makes the solution particularly appealing for confidential discussions, such as in executive boards, finance departments, or negotiations with business partners. Plugins or integrated security features for video conferencing systems like Teams or Zoom are also conceivable. Alternatively, a protected corporate infrastructure could centrally analyze particularly important conversations.

Next steps: Practical Testing and Data Protection

The demonstrator is currently still in the proof-of-concept phase. In the next step, the researchers aim to collaborate with companies and providers of video conferencing systems to integrate the technology in a practical, privacy-compliant, and user-friendly manner. Legal questions also need to be clarified, such as whether conversation partners must consent to the analysis.
A technical detection alone, however, is not sufficient. In cases of unusual demands—such as a request for a sudden transfer—companies should always verify the identity of the counterpart through an independent communication channel. Follow-up questions, a callback, or a predefined approval process can prevent a convincing deepfake from becoming a costly fraud case.

Complement Technical Detection with Clear Processes

A technical detection alone is not sufficient. In cases of unusual demands—such as a sudden request for a transfer—companies should always verify the identity of the counterpart through an independent communication channel.Follow-up questions, a callback, or a predefined approval process can prevent a convincing deepfake from leading to a costly fraud case.