Paper Title
STUDY OF PAYMENT FRAUD DETECTION ON s390x ARCHITECTURE USING IBM’S INTEGRATED ON-CHIP AI ACCELERATOR

Abstract
The COVID-19 pandemic catalysed a rapid shift for contactless online payment transactions. However, the adoption of digital payment systems has led to significant increase in online payment transaction fraud. With AI being an evolving field, it is now evident that detecting fraud has become more effective. Numerous case studies have demonstrated the successful implementation of fraud detection systems on cloud platforms, hybrid environments, and on- premises data centers. The persistent role of IBM Z systems in modern banking is undeniable and have been a cornerstone of banking infrastructure for decades, handling millions of transactions per second and its footprint keep increasing. In such robust system, the application of AI is vital. It is not just about applying algorithms; large banks and payment processors face significant limitations with their existing solutions, allowing them to apply AI to only a fraction of incoming transactions in real-time due to throughput and network latency constraints. The invention of the IBM Z16 Telum processor, with its integrated On-Chip AI accelerator, makes the platform more efficient for AI. In this paper, we present our observations on applying popular AI frameworks on s390x architecture to detect fraud by utilizing various ML/DL models deployed on high performant server like Triton Inference Server. The results show how various AI approaches, including traditional Machine Learning (ML) algorithms like Scikit-learn, accelerated ML algorithms like Snap ML, and Deep Learning (DL) algorithms like LSTM, have been utilized on s390x to achieve high throughput and low latency Keywords - Payment Fraud Detection, s390x, IBM z16, On- Chip AI Accelerator, Triton Inference Server, Machine Learning, Deep Learning, Snap ML, ONNX-MLIR, LSTM, Random Forest, XGBoost, Sci-kit learn, Gradient Boosting, Latency, Throughput