Roskomnadzor became interested in the truth – Newspaper Kommersant No. 204 (7405) dated 02.11.

As Kommersant found out, Roskomnadzor became interested in the development of NRU ITMO in the field of lie detection from video using artificial intelligence. The prototype solution was tested ahead of the US Senate elections. The department admits that the system can be used to detect “deepfakes”. But market participants doubt the prospects of the development, believing that the probability of false positives will be too high.

Specialists from the National Center for Cognitive Research (NTsKR) ITMO are developing an Expert service based on artificial intelligence (AI) for the Main Radio Frequency Center (GRC) subordinate to Roskomnadzor, which will allow checking video recordings of speeches for lies and manipulations. Sources told Kommersant about this and confirmed it in the department. According to one of Kommersant’s interlocutors, the prototype was presented to Roskomnadzor at a meeting of the GRFC AI expert council in July: “The head of Roskomnadzor Andrey Lipov was present, who was interested in the decision. He asked to transfer it to the department for testing as soon as possible.

The service allows you to analyze video recordings for lies and is of interest for the rapid detection of “deepfakes” (deepfake), said the head of the Scientific and Technical Center of the Federal State Unitary Enterprise GRCHTS Alexander Fedotov: “If ITMO can present a sustainable technology in the form of a finished product, GRCHTS will consider the possibility of its application” .

As ITMO explained, the Expert system analyzes video or audio, evaluating confidence, the level of internal and external aggression, congruence (the level of consistency of information transmitted verbally and non-verbally) and inconsistency, and compares words with scientific articles and statements of other specialists. The system is “tailored to work with the English language, but indicators of confidence, aggressiveness, congruence can be extracted in any language.”

Oleg Basov, head of the Cognitive Non-Verbal Research Laboratory of the NCCR ITMO, explained that English was chosen as the most common language: “Improvement of the system for obtaining text from speech in Russian is in the nearest plans.” The system was pre-tested on the eve of the US Senate elections, Oleg Basov said: “We analyzed the interviews of six candidates published on YouTube for inconsistency, aggression and self-confidence.” According to him, the system detected hidden aggression in the statements of three candidates, and one of them was caught lying.

Now more than 90% of Russian developments related to AI are built on the basis of downloaded open foreign libraries along with data sets for their training, explains a top manager of a specialized IT company: “Therefore, ITMO development is adapted for the English language. Retraining for Russian will require a large amount of data and costs.” Similar developments were carried out in the West: in 2017, in the USA, the University of Maryland and Dartmouth College created a neural network that recognizes lies, which were trained on video recordings of trials. The American Innocence Project was also involved in the recognition of lies in court through AI, using Amazon Mechanical Turk for analysis.

Russian developers are already testing solutions for recognizing deepfakes. VisionLabs has expertise in the development of face “deepfake” detectors, and they are effective, they say in the company: the accuracy of the work is 94-100% But now the quality of the neural network is assessed in the plane of probability theory, says the interlocutor of Kommersant in the IT market: “When such a development will even reach pilot operation, the probability of false positives will be quite high, and it will take years for the project to begin to bring benefits. Kirill Lyakhmanov from EBR emphasizes that the very idea of ​​analyzing the “truthfulness” or “falsity” of statements on the basis of indirect signs related to human behavior does not stand up to criticism: “Any kind of “lie detectors” are not recognized as evidence in the courts of most countries.”

Tatyana Isakova, Nikita Korolev, Timofey Kornev