فهرست نشریات علمی دسترسی آزاد ایران

Due to the increase in financial transactions in banking, new technologies and artificial intelligence can help identify cases of money laundering and various types of fraud, thereby reducing many risks. Deep learning and machine learning methods play an effective role in detecting fraudulent transactions and money laundering. One of the main challenges in detecting fraud and money laundering using machine learning and deep learning methods is the imbalance in the training data set, the large volume of samples for learning, and the number of features. In this manuscript, a fraud and money laundering detection approach is presented in two distributed architectures based on the Apache Spark architecture and a centralized architecture. In the proposed Apache Spark architecture, the adversarial generative neural network method is used to balance the dataset. In this approach, three meta-heuristic algorithms are used for feature selection, and classification with embedding learning is performed using three methods: LSTM, MLP, and RF. In the proposed centralized approach, a combined CNN+XGBoost architecture is used, with feature selection based on AO and RTH algorithms. Evaluations show that there is no significant difference in the accuracy index between the distributed and centralized approaches for detecting fraud and money laundering. However, the distributed method in the Apache Spark architecture achieves 8-cluster detection speeds about 7.42 times that of the centralized architecture. In the Apache Spark-based architecture, the proposed method achieves accuracy, sensitivity, and precision of 99.96%, 99.72% and 98.62%. The proposed approach is more accurate than the CNN, XGBoost, RF, MLP, LSTMT, BiLSTM, LSTM-RNN, CNN-BiLSTM, CNN-ELM, and RNN-LSTM (Attention) methods in detecting fraud and money laundering.