Financial Transactional Fraud Detection using a Hybrid BiLSTM with Attention-Based Autoencoder

In this study, we propose an original hybrid model that consists of a Bidirectional LSTM (BiLSTM) and an Attention-Based Convolutional Autoencoder (CAE) designed for fraud detection in financial transactions. The structure of the model is constructed with three Conv1D layers on the CAE and a dense layer that functions as a bottleneck for effectively squeezing relevant information from the transaction data. The importance of certain http transactions can be highlighted using an attention mechanism which helps the model to concentrate on the important features. These features are further fed into the BiLSTM, where the BiLSTM learns to model the context from both past and future sequences of transactions, thus providing a more complete picture of the transactions. To this extent, the model evaluates the reconstruction losses to label the types of fraudulent transaction activity. The performance of this model is found to be very good as it achieved an accuracy of 97% and a high Area Under the Curve in ROC analysis out of the total 100 percent showcasing the model's ability to correctly classify the non-fraudulent and fraudulent transactions.