Predictive analysis of income tax fraud using deep learning techniques

Governments must address income tax fraud as it causes significant revenue loss and threatens financial stability worldwide. Traditional fraud detection systems based on heuristic or rule-based approaches are increasingly ineffective against sophisticated modern fraud techniques. These methods rely on predefined patterns and human intervention, limiting their ability to detect complex anomalies in large financial datasets. This study proposes a deep learning-based approach using a hybrid model that combines Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and Autoencoders (AE). The model captures spatial and temporal patterns and extracts hidden features for accurate anomaly detection, and is evaluated on over 8,500 financial records. The model achieves 97.52% accuracy, with 95.28% precision, 94.85% recall, an F1-score of 95.10%, and an AUC-ROC of 98.36%, outperforming traditional methods. These results demonstrate that deep learning offers a robust, scalable, and automated solution for detecting tax fraud and improving compliance. This performance supports the objective of accurately identifying fraudulent tax filings in complex financial datasets. The integration of CNN, LSTM, and autoencoder models enables effective detection of spatial, temporal, and behavioural fraud patterns.