Fraud Detection using IBMs Differential Privacy Library

Abstract

Abstract: In the year 2022, a total of 602 billion rupees were lost on bank frauds consisting of over 9103 major fraudulent cases. Due to the increasing number of frauds all around the world, it is vital to safeguard consumers' privacy. There is a prediction of a spike in cyber credit fraud in the future years. There are currently various techniques for credit card fraud detection but all the ML models have lower accuracies and cannot cope with the real-time datasets as the transaction data is extremely confidential. To address this issue we have used differential privacy for fraud detection. Differential privacy helps researchers to gain sensitive information about individuals without compromising their privacy. IBM provides a library: Diffprivlib, for exploring, researching and developing applications in Differential Privacy. In this paper, we have experimented with the impact of differential privacy on a sample dataset using various machine-learning models. With a comparative study of the algorithms, we propose that Isolation Forest has the highest accuracy. It can be added to the diffprivlib library since the library is open source and Isolation Forest does not exist in the library till date. Furthermore, we have built a system to predict fraudulent transactions.