Roskomnadzor launched the Oculus automatic search system for prohibited content. This was told to Vedomosti by a representative of the Main Radio Frequency Center subordinate to Roskomnadzor (FSUE GRCHTS), who acted as the customer for the project.
“The Oculus information system has already been launched and is fulfilling its tasks in full: it detects violations of the law in images and video materials,” a representative of the GRCHTs told Vedomosti. According to him, in December 2022 the system was tested, and in January 2023 the integration of the system with other Roskomnadzor monitoring tools began. The representative of the GRFC did not give details of the test results, as well as the first results of her work.
The main task of the system is to identify violations of Russian legislation in images and videos, the representative of the GRFC specified. “The system recognizes images and symbols, illegal scenes and actions, analyzes text in photo and video materials. Oculus automatically detects such offenses as extremist themes, calls for mass illegal events, suicide, pro-drug content, LGBT propaganda, etc.,” he explained.
Oculus is a system that involves performing tasks for classifying images and videos according to specified requirements, including the main types of prohibited content, Natalia Tylevich, general director of the Social Laboratory company, explained to Vedomosti. According to her, “Oculus” works as a classifier with a predefined set of information sources that analyze content for compliance or non-compliance with the requirements of the law, she added. That is, Oculus can analyze specific pages of sites or publics and profiles in social networks. The program does not collect data, it classifies them.
Prior to the introduction of the system, GRCHTS specialists analyzed prohibited content “mostly manually,” said a representative of a body subordinate to Roskomnadzor. “On average, operators processed 106 images and 101 videos per day. Oculus will analyze more than 200,000 images per day (about three seconds per image). The system will automate and significantly speed up the monitoring of visual content,” continued the representative of the GRFC. This, officials hope, will “improve the effectiveness of identifying signs of violations.”
In the future, until 2025, the system is planned to be improved. “The possibility of adding new classes and types of violations, as well as the function of determining the postures of people and their actions, is being worked out,” the representative said. The system will detect prohibited material “on multiple frames of video clips, in complex handwriting and hand-drawn content,” he added.
For the first time, information about the development of the Oculus system appeared in mid-2021, when the GRFC published a procurement for the development of technical specifications for the system worth 15 million rubles. In the future, the GRFC also published a purchase directly for the development of Oculus itself – already for 57.7 million rubles. At the same time, the total cost of a set of solutions that will effectively find and block various types of prohibited content is estimated at about 1.5 billion rubles, says Vedomosti’s interlocutor in one of the domestic developers of IT solutions. All this may indicate that the development of the GRFC and involved companies on this toolkit will continue, he added.
The developer of Oculus was the IT company Execution RDC. Its representative declined to comment, citing commercial relations with the customer.
The need to use an automated solution to search for prohibited content in the GRFC is explained by the aggressively growing flow of such materials on the Internet. Especially a lot of recently began to appear prohibited materials related to the theme of the military operation on the territory of Ukraine. We are talking about “unprecedented volumes and speed of distribution of fakes, which are aimed at replacing real facts with a specially constructed” reality “,” the representative of the GRFC notes.
According to him, in 2022, on the basis of the requirements of the Prosecutor General’s Office, more than 100,000 Internet resources were deleted or blocked, which contain false information (including about the course of a military special operation). In 2021, there were about 7,000 such materials, in 2020 – about 1,500, and in 2019, several hundred. “The most used tool in the hands of anti-Russian sources has become visual content, which is able to make a greater impression on the emotions of users,” added the representative of the GRFC.
But in addition to the topic of the military operation, an array of materials promoting drug use, calls for suicide, child pornography is also growing on the Internet, we are talking about “an avalanche-like growth of stuffing in all types of prohibited information,” the representative of the GRFC added. “Such content needs to be found and blocked before it is widely distributed online, especially among children and young people,” he stressed.
The amount of destructive content has been growing exponentially since 2017, especially over the past year, the distribution of various extremist and terrorist materials has increased, including calls for sabotage aimed at teenage and youth audiences, confirms Ekaterina Mizulina, CEO of the Safe Internet League. There is also a growing number of banned LGBT and drug-related material, she added.
“Indeed, artificial intelligence systems for the analytics of such information are very necessary. There are a number of independent systems, including those that we use in our work, but there was no such state system until recently,” Mizulina noted. The risks of using Oculus relate to how correctly or incorrectly the classification of images and videos is carried out and how correctly they will be further interpreted in conjunction with other components of the system and a person, Tylevich continues.
These risks are due to such limitations of artificial intelligence systems, such as, for example, a false positive, when a particular content is determined to be prohibited, or, on the contrary, a pass, when the system does not notice images or videos that actually have the characteristics of prohibited content. “The issue of interpretation and settings will determine the scope of such errors, how controversial content will be classified – whether to consider it prohibited or not. But one way or another, these will be the risks not so much of the decision itself, but the risks of further work with the data obtained as a result of its work,” the expert concluded.
“Of course, after artificial intelligence analytics, it is very important to manually select such information, check the work of AI and already work with specific facts,” Mizulina agrees with her. Obviously, all the revealed information will not be automatically sent anywhere, including for initiating administrative and criminal cases, as well as for entering into the register of prohibited sites. “For now, it would be premature. Therefore, at the next stage, after the work of artificial intelligence, standard, manual data processing is also important,” summed up the CEO of the Safe Internet League.