Oculus System Will Search for Illegal Content on Runet

According to media reports, Roskomnadzor has selected a developer for the “Oculus” system, which will search for illegal content online. The project is allocated 57.7 million rubles, and the neural network-based system will analyze photos, videos, and texts on websites, social networks, and messengers in real time for prohibited information, including the promotion of homosexuality, drug and weapon manufacturing. The Oculus system is expected to be launched by December.

The publication Kommersant, citing the results of a government procurement tender, reports that the development of the Oculus system will be handled by LLC “Execution RDC.” The company will receive 57.7 million rubles for the project, which must be completed by December 12 of this year.

Journalists note that experts doubt whether the allocated funds are sufficient for development and whether the contractor can meet the deadline. For example, creating training datasets for the neural network alone would require much more time.

According to the tender terms, the contractor must not only create the system but also provide the “Main Radio Frequency Center” (FSUE GRChTs, a Roskomnadzor structure) with datasets that will help identify prohibited information. The contractor will also supply GRChTs with labeled data sets for training a facial recognition model, with at least 100 samples, since Oculus must include functionality for cataloging identifiable symbols, scenes, actions, and individuals. It is specifically emphasized that the system should detect illegal content not only in text but also in “scenes, object combinations, image compositions, faces, and both static and dynamic movements.”

The system’s capacity should be 200,000 images per day, meaning Oculus must analyze two frames or images every second.

According to Konstantin Bulanov, Director of Digital Technologies at GRChTs, the system will require at least 48 servers equipped with graphics accelerators for stable operation. The search for illegal content will be performed by neural networks based on deep machine learning.

Bulanov explained to RBC that Oculus does not independently search for or collect information on the internet: “Other information systems, which are already part of the Unified Analysis Module (EMA), are used for monitoring. The Oculus system will also be integrated into EMA. Data from monitoring systems will be automatically uploaded to Oculus. The new system will assess whether the material contains signs of information prohibited by law. If such signs are detected, an automatic report will be sent to operators for expert confirmation and further action.”

A large unnamed IT company familiar with the project’s technical requirements told Kommersant journalists that, in reality, implementing such a system is nearly impossible—not only within such a short timeframe and for 57 million rubles, but in principle: “The contractor is not given specific goals, tasks, and requirements that could be met by refining an existing system, but is instead asked to design a fundamentally new solution, which cannot be done in the specified time.”

At the same time, Vladimir Arlazarov, head of Smart Engines, believes that “at the current level of IT development,” the stated tasks can be accomplished with about 90% quality, but he also expressed surprise at the tight deadlines. According to him, “just collecting representative training datasets could take much longer.”