Anomaly Detection with Variational Autoencoders

Credit card fraud has emerged as a pervasive threat, impacting financial institutions and individuals as online banking and payment methods become increasingly integral to daily life. Despite efforts to mitigate this problem through measures like passwords and two-factor authentication, financial institutions continue to suffer substantial losses, often amounting to millions of dollars. Traditional machine learning solutions, developed and trained as supervised learning models, have failed to address this issue effectively. In anomaly detection, such as credit card fraud detection, the available training datasets are vast but inherently imbalanced, posing a formidable obstacle for supervised learning models in developing accurate hypothesis functions during model training. In response to this challenge, we propose an alternative approach using unsupervised machine learning models, specifically the Variational Autoencoder. In our study, we constructed the Encoder and Decoder components of our VAE using deep neural networks, implemented and trained with pre-processed data using the TensorFlow and Keras libraries. Our model's effectiveness was assessed using the AUC-PR score, yielding a robust score of 88%. Additionally, we subjected our model to an accuracy test on a test dataset, achieving a commendable accuracy score of 78%. The results of our study underscore the viability of unsupervised learning models, particularly VAE, for credit card fraud detection, demonstrating their potential as a machine learning solution to mitigate credit card fraud. This research offers a promising avenue for bolstering security measures in financial transactions and reducing the impact of fraudulent activities.