Privacy-Preserving Machine Learning for Secure Business Intelligence

Abstract

Business intelligence (BI) increasingly relies on machine-learning models trained on customer, employee and transaction data, and these models can leak the data they were trained on through their parameters and outputs. Privacy-preserving machine learning offers formal protection, principally through differential privacy (DP), which bounds what any output reveals about any individual record, and through federated learning, which keeps data at its source. Both impose a utility cost. This paper examines the privacy–utility trade-off for a representative BI regression model in a simulation of 4,000 records. Without protection the model attains a hold-out R² of 0.69. Centralised training with output-perturbation DP retains R² of 0.68 at a privacy budget of ε = 0.5 and 0.64 at ε = 0.25, falling to 0.38 at ε = 0.1; federated training with per-client DP across five sites loses utility faster, retaining 0.62 at ε = 0.5 and 0.42 at ε = 0.25. The paper explains why local protection costs more than central protection, maps the trade-off to BI use cases by sensitivity and precision requirements, and proposes an architecture and governance for privacy-preserving BI in which the privacy budget is a managed resource. The contribution is a practical account of how organisations can obtain analytical value from sensitive data under formal privacy guarantees.