Scalable financial anomaly detection implementation in a distributed graph database server

Abstract

Financial transactions have become a prominent part of the economy in the world. Over 1 billion transactions are being processed daily around the world and a considerable portion of those transactions accounts for various fraudulent activities that take place around the world. Terrorist Financing, Money Laundering are popular examples that generate fraudulent transactions. Financial institutions are obligated to have the capability to detect such transactions and perform necessary measures to mitigate and report the parties involved with such fraudulent transactions. Several implementations exist that map the financial transactions in the form linked/graph data and detect any anomalies using the structural features of the graph such as PageRank, Degree Distribution, etc. However, these implementations require the transactions to be executed in a single graph database and this limits the capability to scale horizontally when the number of transactions increases. This research proposes an extension to a C/C++ based distributed graph database server called JasmineGraph that is capable of handling large amounts of graph data and performing anomaly detection algorithms in a distributed manner. We generate Degree Distribution and PageRank scores for the graph network in a distributed manner and use these graph structural features to train a machine learning model for anomaly detection. Our distributed anomaly detection approach has been able to predict anomalous transactions with an F1-score up to 0.98 and was able to reduce the execution time by 79.5% in comparison to the non distributed approach when detecting anomalies. For large datasets (PaySim-2M), the non distributed approach failed to process due to lack of memory but was successful after using the distributed approach making it more efficient to use our distributed anomaly detection for large financial transaction networks. As future work, we plan to expand our anomaly detection approach on streaming graphs for real time anomaly detection.