The Use of Machine Learning Techniques to Track Social Engineering in Banking Transactions

The relevance of the study is due to the fact that in the era of digitalization and a widespread introduction of technologies, online payments and other remote banking transactions are becoming increasingly popular, which leads to an increase in cases of fraud using social engineering. The purpose of the study is to assess the opportunities of machine learning methods to counteract social engineering, as well as to identify limitations of using these methods in banking anti–fraud systems. Materials and methods. To test the hypothesis of insufficient effectiveness of machine learning methods for recognizing attacks using social engineering, the Bank Account Fraud Dataset Suite (NeurIPS 2022) was used. Data preparation for machine learning was implemented using UMAP nonlinear dimensionality reduction algorithm and SMOTE synthetic oversampling method. Study results. Clustering of data did not make it possible to reliably separate fraudulent and non-fraudulent transactions depending on any characteristics of the client or the transaction itself. At this, the accuracy of the model based on decision tree is the highest compared to gradient boosting and logistic regression, however, the efficiency and effectiveness of practical using the algorithms inevitably decreases due to the drift of concepts, which causes an increase in the probability of erroneous classification. In case of concept drift, the model becomes over-trained and does not work well on test data. An early stop can be used for regularization, when the model learning process should stop at the point where losses and errors in the verification data reach a minimum value. Conclusions. Since banks are obliged to counteract transfers that occur without the client's voluntary consent, in particular under the influence of intruders, the anti-fraud systems of sending and receiving banks should be regularly retrained in order to avoid the spread of fraudulent cases and banks receiving losses from the need to make compensation payments. The practical significance lies in the possibility of using the results to track and counteract social engineering in banking transactions in order to further improve regulatory efforts and actions of commercial banks themselves.