Multiclass Detection of e-Wallet Fraud Transactions Using Deep Learning Techniques

BROU Pacôme, KOUASSI Francis, KOUASSI Thomas and ASSEU Olivier

The widespread adoption of digital payments through Electronic Wallets (e-Wallets) has significantly increased the exposure of financial systems to sophisticated fraud schemes and abnormal transactional behaviors. This situation raises a critical question: How can abnormal, potentially fraudulent, transactional behaviors be detected in real time and with high reliability within massive, sequential streams of heterogeneous data? To address this challenge, this study proposes a hybrid approach combining a Long Short-Term Memory (LSTM) recurrent neural network capable of capturing the temporal dimension of user behaviors with a multinomial logistic regression (MLR) classification layer to discriminate between behavioral classes. Using a simulated dataset of 1,000 transactions from 100 users, where each transaction was enriched with contextual variables (device_score, frequency_score, timestamp, amount, location, transaction_type), the model classified behaviors into three categories: Normal, Suspicious and Fraudulent. The hybrid model demonstrated strong overall performance, achieving an average accuracy of 77.3%. It exhibited excellent recall for the Normal class (91%), acceptable performance on Suspicious transactions (73% recall) and a robust ability to detect fraud (76% recall), while reducing false positives by 35% compared to a standalone static classification. The temporal integration enabled by the LSTM significantly improved the detection of gradual behavioral drifts, particularly in cases where fraud leveraged historically trustworthy devices. This work highlights the value of a sequential and adaptive approach to enhancing transactional cybersecurity in environments characterized by high behavioral variability.

View Fulltext Back