Employees and students of the Department of Cryptology and Cybersecurity of National Research Nuclear University MEPhI have created the "Sipuha" system, which is capable of identifying fake audio recordings (voice deepfakes). The project, developed under the leadership of Associate Professor Dmitry Efanov, won the "Innovator of Moscow" competition.
The system is based on a neural network trained on more than 200 thousand audio recordings, including both real voices and deepfakes generated by different algorithms. The analysis is carried out using cepstral coefficients — mathematical characteristics of sound, calculated for short fragments of the recording.
The development began in 2022 as a scientific project with the support of the Foundation for Assistance to Innovations. By 2024, the technology became in demand due to the increase in fraud with fake voices. "Sipuha" was among the finalists of the MEPhI and Rosatom accelerator.
The team is currently working on a cloud service to integrate the system into call centers and messengers. TBank and Informzashchita have shown interest in the project. In the future, it is planned to create a laboratory for continuous improvement of algorithms. The developers are counting on the support of the Government of Moscow.
Read more materials on the topic:
Neural networks show high uniqueness: more than 80% of generated texts pass the test
Made by AI: it was proposed to introduce labeling of products generated by neural networks in Russia
