A deployment-derived online banking fraud detection inference-log dataset from a live cloud-based deep learning...

This article describes a dataset of 56,962 transaction inference logs collected through a live, cloud-deployed online banking fraud detection demonstration system over a 31-day collection period from 01 January 2026 to 31 January 2026. The data were gathered through a REST API endpoint hosted on a Ubuntu Linux virtual private server at model-test.online, where a trained hybrid Convolutional Neural Network and Long Short-Term Memory (CNN-LSTM) model processed each incoming transaction in real time. The dataset contains 98 confirmed fraudulent transactions (0.172% fraud rate), reflecting the natural class imbalance encountered in operational environments. Each record comprises 38 columns in total: 30 input features consisting of 28 principal component analysis (PCA)-transformed variables (V1 through V28), the transaction amount (in USD), and a time index; and 8 additional fields including model-generated outputs (fraud probability score, categorical risk level [LOW, MEDIUM, or HIGH], confidence estimate, and action recommendation [BLOCK or ALLOW]), response latency in milliseconds, a UTC timestamp, a partially anonymised IP address, and a transaction identifier. Ground truth fraud labels were assigned through manual verification of model-flagged (BLOCK) transactions during the deployment period; transactions that received an ALLOW recommendation were not individually reviewed and were assigned a legitimate label by default. The deployment was a proof-of-concept system operating as a public demonstration interface and not a licensed banking institution; accordingly, the transactions represent API submissions by the public and the authors during testing, sharing the PCA-transformed feature structure of the ULB Credit Card Fraud Detection benchmark. The dataset is freely available on Zenodo (DOI: 10.5281/zenodo.20030064, CC BY 4.0) to support reproducibility of the accompanying research, to provide a recent deployment-derived fraud benchmark, and to facilitate independent evaluation of fraud detection methods under realistic class imbalance conditions.