An enhanced LSTM architecture for financial fraud detection with temporal attention, intermediary memory, and camouflage detection

Abstract

Financial fraud detection remains a critical challenge due to the dynamic, imbalanced, and often camouflaged nature of fraudulent transactions. Traditional recurrent models, such as LSTM and GRU, capture sequential dependencies but fail to prioritize temporally relevant events or model complex inter-transaction relationships. More advanced variants, such as Time-LSTM and Phased-LSTM, address irregular temporal patterns but lack mechanisms to detect collusive or camouflaged behaviors. In this paper, we propose a novel enhanced LSTM architecture that integrates four key components: a Temporal Attention Gate to dynamically emphasize recent and contextually significant transactions, an Intermediary Memory Gate to capture cross-transaction dependencies, a Camouflage Detection Gate to identify similarity-based collusion patterns, and Conv1D layers for efficient local feature extraction. The proposed model is evaluated on two real-world fraud detection datasets (Credit Card Transactions and IEEE-CIS) under multiple class imbalance scenarios (5:1, 15:1, 27.3:1, 100:1, and 164.6:1). Experimental results demonstrate that our method consistently outperforms LSTM, GRU, Time- LSTM, and Phased-LSTM across all metrics (Precision, Recall, F1-Score, and AUC). Notably, the proposed method exhibits superior robustness, with only 7.6% points average degradation from mild to severe imbalance on the IEEE-CIS dataset, compared to 12.8% points for the baselines. The performance gap widens as the imbalance increases, achieving up to 32.8% relative improvement in AUC over the best baseline on the Credit Card dataset (at 100:1 and 164.6:1 ratios) and 19.1% on the IEEE-CIS dataset (at 27.3:1 ratio). These findings confirm that the proposed architecture effectively addresses temporal prioritization, relational modeling, and camouflage detection, making it highly suitable for real-world fraud detection in highly imbalanced financial environments.