MTS Web Services: Deepfake Detector

Main article: Deepfakes (DeepFake)

2025: Deepfake Detector Launch

MTS PAO, a digital ecosystem, announced on November 21, 2025 that its subsidiary had MTS Web Services (MWS) launched a detector. deepfakes The system recognizes with more than 98% accuracy, content created by - AI models such as Veo 3 () Google and Sora 2 (), OpenAI which are able to generate and edit videos according to text description.

To recognize deepfakes, audio track analysis using a detector from MWS AI and image and video analysis using a detector from VisionLabs are used. The sound part uses a specialized model: it first learns from raw recordings of human speech in order to "understand" natural sound patterns, and then is further trained on synthetic recordings, revealing the characteristic features of the generated voice.

The high accuracy of recognition of videos created by AI models allows you to automatically identify fake videos before publication, filter AI content in media, social networks and instant messengers, prevent the spread of videos with fake images of politicians and public figures, and counteract attempts at mass manipulation, for example, when creating fake video disasters and emergencies that can cause panic.

Audio track recognition accuracy was 84% for Veo 3 videos and 93% for Sora 2. When analyzing the image, the accuracy of the detector from VisionLabs reached 93.9% for Veo 3 and 93.6% for Sora 2. At the moment, training of algorithms is ongoing to achieve a target accuracy above 98%, as well as combining audio, image and video recognition technologies into a single detector with a common interface.

{{quote "Sooner or later, new types of synthesis appear that bypass existing detectors. From this point of view, our main task is to regularly update neural networks and do it as quickly as possible, so we focused on the speed of additional training of algorithms. With the Veo 3 and Sora 2 models, we were able to adapt in just two weeks - and this is our main advantage now, - commented Pavel Voronin, General Director of MTS Web Services. }}

One of the key difficulties when working with generative models is the increase in the False Rejection Rate (FRR) - the number of false deviations when the system takes the real voice for the synthesized one. This can happen when using sound enhancement tools such as noise cancellation, compression, or voice filters. According to preliminary estimates, FRR when analyzing such records can reach from 4.5% to 7.2%, which requires additional adjustment of algorithms.

As of November 2025, MWS is testing a new deepfake detector with a platform for video conferencing and online training MTS Link, in the MTS Defender service, which warns users about a conversation with a possible fraudster, with one of the state services in Russia, as well as three banks in Russia and the CIS.