VAE-TF: A Variational Autoencoder Framework for Financial Fraud Detection

In the context of modern digital transactions, the sheer volume and velocity of transactions necessitate robust financial fraud detection mechanisms to effectively distinguish between fraudulent and legitimate activities. A key weakness of traditional models is that these models are susceptible to performance degradation in presence of imbalanced data distributions, which leads to poor generalization performance. To address these limitations, we propose VAE-TF, a Variational Autoencoder-based fraud detection framework that learns the underlying distribution of normal transactions and identifies fraud as deviations from this learned representation. The proposed method consists of comprehensive data preprocessing, including handling missing values, feature scaling using StandardScaler and dimensionality reduction via Principal Component Analysis (PCA) to enhance feature representations. To mitigate class imbalance, we employ the Synthetic Minority Over-sampling Technique (SMOTE) and an unsupervised learning paradigm where VAE is trained solely on normal transactions. The VAE architecture consists of an encoder that maps input transactions into a latent space by learning a probabilistic distribution parameterized by a mean and variance, followed by a decoder that reconstructs transactions from the latent space. A combined loss function, assessing the precision of transaction reconstruction, and a KL divergence loss, facilitating latent space regularization, guides the framework’s training. Instead of using a fixed threshold for anomaly detection, we propose a ranking-based fraud identification approach, where the top k transactions with the highest reconstruction errors are flagged as fraudulent. The proposed framework is evaluated on multiple financial fraud datasets, demonstrating its superior performance in detecting fraudulent transactions. The results indicate that VAE-TF is a highly effective framework for anomaly detection in financial fraud scenarios.