A Deep Learning Model for Investment Scam Prevention in Financial Systems: Autoencoder-based Anomaly Detection

Financial and investment scams have become a significant threat to both individual investors and financial institutions. Detecting fraudulent transactions remains a challenging problem due to the constantly evolving tactics employed by scammers. Current methods primarily rely on predefined rules or expert knowledge, which can easily be bypassed by sophisticated fraud schemes. Additionally, traditional fraud detection systems struggle with the imbalanced nature of fraud detection datasets, where the number of legitimate transactions far exceeds fraudulent ones, leading to poor generalization and high false positive rates. This paper proposes an innovative approach to tackle financial and investment scams using autoencoders, a type of unsupervised deep learning model. Methodology focuses on anomaly detection, allowing the model to learn normal transaction patterns and flag anomalies indicative of fraud. Enhancing this approach by employing Z-score standardization for normalization, ensuring that all features contribute equally to the model's learning process. Mean imputation is used to handle missing data, preserving the integrity of the dataset. The autoencoder architecture incorporates ReLU activation functions in the encoder and decoder layers, allowing the model to learn complex patterns in the data, while the output layer uses the sigmoid function to provide reconstructed values within a normalized range. By training exclusively on normal transactions, the model learns to accurately reconstruct legitimate behavior and flag outliers as potentially fraudulent. The approach offers a significant advantage over traditional methods by being capable of detecting novel and previously unseen fraud schemes, making it an invaluable tool for real-time fraud detection and monitoring.