Neural networks used by banks to protect customers may miss fraudulent transfers not due to weak algorithms, but due to a lack of real examples for training. This conclusion was reached by researchers from Moscow Polytechnic University.

Almost all banking operations are legitimate, while the share of thefts is only thousandths of a percent. As a result, the model sees millions of ordinary payments and isolated cases of fraud. The easiest way for it to reduce the number of errors is to almost always respond that the operation is safe. Formal accuracy remains high, but fraudsters pass the check.

Developers are trying to correct the imbalance using synthetic theft cases, increased penalties for missed attacks, and combinations of several algorithms. Graph neural networks additionally analyze the connections between accounts and operations, so they can find not a single suspicious transfer, but a whole group of related participants.

The best machine learning systems show accuracy from 88 to 94%, but quickly become outdated due to the emergence of new fraud schemes. Banks, however, cannot freely exchange real cases due to customer confidentiality, and complex models are often unable to clearly explain why a specific operation was blocked.

According to the Bank of Russia, in 2025, fraudsters stole 29.3 billion rubles from Russians – 6.4% more than the previous year. Less than 6% of the stolen amount was returned to customers.