DeepFinSec: Advancing Online Financial Fraud Detection via Structural Similarity-Based Graph Embedding

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The frequent use of digital technology has significantly increased the use of online transactions. Most financial activities are conducted electronically. Financial fraud involves deceiving individuals or businesses for monetary gain. The proposed graph node embedding technique, FA-Struc2vec, is applied to the financial transaction network to extract graph node features based on their structural similarity and generate node embeddings. The node embeddings serve as input to DeepFinSec, which classifies transactions into desired classes. DeepFinSec is a proposed DL-based model that is a hybrid of CNN and LSTM. CNN processes transaction data and extracts local and high-level abstract features while LSTM captures sequential and temporal dependencies, making this a powerful combination for transaction pattern analysis. DeepFinSec outperformed existing ML and DL techniques, achieving 97.70% accuracy, 2.29% higher than the second-highest-accuracy technique, LSTM. 95.61% precision, 100% recall, 97.75% F1 score, and 99.68% AUC.