Enhanced autoencoder-based fraud detection: a novel approach with noise factor encoding and SMOTE

Abstract

Fraud detection is a critical task across various domains, requiring accurate identification of fraudulent activities within vast arrays of transactional data. The significant challenges in effectively detecting fraud stem from the inherent class imbalance between normal and fraudulent instances. To address this issue, we propose a novel approach that combines autoencoder-based noise factor encoding (NFE) with the synthetic minority oversampling technique (SMOTE). Our study evaluates the efficacy of this approach using three datasets with severe class imbalance. We compare three autoencoder variants—autoencoder (AE), variational autoencoder (VAE), and contractive autoencoder (CAE)—enhanced by the NFE technique. This technique involves training autoencoder models on real fraud data with an added noise factor during the encoding process, followed by combining this altered data with genuine fraud data. Subsequently, SMOTE is employed for oversampling. Through extensive experimentation, we assess various evaluation metrics. Our results demonstrate the superiority of the autoencoder-based NFE approach over the use of traditional oversampling methods like SMOTE alone. Specifically, the AE–NFE method outperforms other techniques in most cases, although the VAE–NFE and CAE–NFE methods also exhibit promising results in specific scenarios. This study highlights the effectiveness of leveraging autoencoder-based NFE and SMOTE for fraud detection. By addressing class imbalance and enhancing the performance of fraud detection models, our approach enables more accurate identification and prevention of fraudulent activities in real-world applications.