FraudAmmo: Large Scale Synthetic Transactional Dataset for Payment Fraud Detection

Global losses due to payment fraud have tripled from $9.84 billion in 2011 to $32.39 billion in 2020 and are expected to reach $40.62 billion by 2027. In addition to the financial losses, fraud negatively impacts brand reputation and leads to a bad customer experience. Advanced machine learning has been actively adopted to tackle the fraud detection problem at scale. However, the scarcity of open datasets leads to less reproducible research, especially in the payments domain. We have released a synthetic transactional dataset, FraudAmmo, containing 3 million transactions, synthetically generated from real-world datasets. FraudAmmo is diverse with respect to transactional channels, geography, customer, and fraud types. We leverage the idea of privacy preservation in tabular Generative Adversarial Neural Networks (GANs) to generate FraudAmmo. In addition to privacy-preserved GANs, we have applied certain checks to ensure the customer's differential privacy. The quality of the generated dataset is evaluated on various metrics including machine learning efficacy, statistical similarity, and privacy preservability. We have also benchmarked our results on FraudAmmo using machine learning algorithms such as bagging, boosting, MLP, and logistic regression. To the best of our knowledge, this is the first large-scale synthetic fraud transaction dataset that aims to aid academia and research groups in developing and validating their fraud detection models. The dataset can be found here - https://github.com/karthi2107/FraudAmmo