SPbPU: Graph neural network to combat fraud

Main article: Fraud Detection System

2023: Announcement of the creation of a graph neural network model

A group of scientists from the Institute of Cybersecurity and Information Protection of SPBPU State University has created a model of a graph neural network that can distinguish suspicious transactions from safe, scammers from honest users. During experimental tests, the model showed its potential. A scientific article with the results of the study conducted within the framework of the Priority-2030 program was published in SpringerLink. This was announced on May 2, 2023 by representatives of SPBPU University.

SPBPUS scientists have taught a neural network to fight fraud. Illustration: innovations.ru.

As reported, graphs are data structures that are networks with paired connections within. As a rule, they are represented in the form of nodes and lines, which are also called edges. And graph neural networks are a type of neural networks that are oriented to work with the structure of the graph. Many data can be represented in the form of graphs - relations between users of social networks, structural representations of proteins and organic compounds, a data transmission network, transactions between bank accounts. Graph neural networks, in fact, combined all the developments in the field of neural networks, which were aimed at processing data represented in the form of objects and relations between them.

About a year ago, Polytech scientists began studying graph neural networks in the banking sector. They analyzed and processed a large amount of data: hundreds of transactions and detailed information about them, starting with the operation number and ending with the type of device with which the translation was carried out. Then the researchers built a model of the graph neural network and moved on to learning it.

We presented banking transactions and users who make them in the form of graphs, then divided them into two classes: some are fraudsters, others are people who make legitimate money transfers. When training our graph neural network, we additionally took into account the identification information: bank card number, data on the sender and recipient of funds, the type of bank card used, the characteristics of the device with which the transaction was made, and others. The identification of additional features allowed us to more accurately train the graph neural network and get good results. shared by Darya Lavrova, PhD, Professor, Institute of Cybersecurity and Information Protection SPBPU

The main weapon of this model of the neural network is that it pays attention to certain patterns by which illegal actions can be recognized. When "filtering" transactions, for example, the neural network looks at the time stamps by which it determines how long ago a person became a member of the banking environment and in which organization he is served.

If a person opened an account with a bank six months ago and during this period of time the average amount of transactions per day was 1,000 rubles, after which on one day he received money transfers in the amount of 30,000 rubles, the likelihood that the neural network will classify this person as a class of fraudsters will increase. In addition, information about the source of the transaction will be taken into account and, if the money was transferred not by a legal organization, but by 10 individuals, then this probability will also increase. gave an example of Darya Lavrova

This neural network model, developed at the Polytechnic, is focused on solving the problem of large amounts of transaction data, optimizing the speed of analyzing operations for safety and detecting other methods bank fraud. According to scientists, for 2023, when each person can make several purchases on the Internet in a day, it is important that there is a protection mechanism that will quickly reveal fraudulent actions from a huge number of banking operations.

The development of the Polytechnic can be very useful primarily for banking organizations. Firstly, it is able to save human resources by automating all the routine work on manual analysis of transactions, which is carried out by bank employees. They will have to deal only with those operations that the neural network considered suspicious. Secondly, banks can save - working with a neural network, organizations will not have to spend the budget on reconfiguring the network infrastructure, purchasing information security tools, training employees in the rules of the so-called "digital hygiene," and most importantly, compensation for damage from fraudsters.

{{quote 'author
= shared Anastasia Sergeeva, one of the authors of the project, a researcher at the Institute of Cybersecurity and Information Protection SPBPU' Our method may well be used as the first line of defense: to reduce the amount of data with transactions and detect many types of fraud. However, our method, like all other technical methods of ensuring information security, will never become the main method of protection simply because the most vulnerable link is not a computer, but a person. As long as users enter their credit card data on third-party sites, do not use reliable passwords and believe calls from alleged bank security officers, security will not be ensured. Development of technical means of protection shall take place in parallel to teaching users the basics of digital literacy and safe behavior on the Internet.}}

In practice, the developers say, the neural network must for some time "get used to" the features of the information environment in which it will work. Therefore, it is better to train her on the data of the same banking environment in which she will detect fraudsters. At the same time, for more accurate operation of the graph neural network, it is better to describe the graph itself in as much detail as possible, that is, to give neural networks more information about users, as well as examples of fraudulent transactions that have already been noticed in the target banking environment.

{{quote 'author
= summed up Dmitry Zegzhda, Doctor of Technical Sciences, Director of the Institute of Cybersecurity and Information Protection of St. Petersburg State University Corresponding Member of the Russian Academy of Sciences' Cybersecurity Assurance - a continuous process, an eternal "arms race" between technically qualified security specialists and violators. Therefore, any protection system, most likely, sooner or later, will be hacked, but other protective mechanisms will be created to replace it. We work towards optimizing our graph neural network model by collecting and generating various training datasets, which will include more "cunning" transactions.}}

The graph neural network, according to Polytech scientists, can be used in different fields, where the data can be represented in the form of a set of objects and the connections between them. So, for example, it will cope with identifying users on social networks who spread misinformation or with the detection of network attacks in data networks.