Privacy-Preserving Federated Analytics Framework for Secure OLAP-Driven Data Mining in Enterprise Data Warehouses
Modern organizational analytics rely on enterprise data warehouses (EDWs). However, large-scale and centralized AI-driven mining of sensitive data stored in these warehouses means that an organization is vulnerable to privacy leaks, inference attacks, and not complying with regulations. Many of the current privacy-preserving frameworks are not designed for direct integration with OLAP environments. As a result, they cannot provide both high analytical utility and formal privacy protection. The framework proposed in this paper provides a definitive solution through integrating OLAP-based multidimensional feature engineering and federated learning (FL) through FedAvg and differentially private stochastic gradient descent (DP-SGD). The system is designed to partition EDW data across multiple logical clients and allows for collaborative training of a single global neural network while keeping all raw data stored locally within each client. Accuracy of classification results for the UCI Adult Income public benchmark and a synthetic EDW-based dataset tested against four baseline models are reported to be 94.8% with a ε of 2.0 for the synthetic EDW dataset, which is only 1.4% lower than centralized. Furthermore, the use of warehouse-aware federated analytics methods has been shown to reduce membership inference attack (MIA) success rates from 79.4% to 54.8%. Finally, the overall system was maintained with a formal (ε, δ)-differential privacy guarantee, thus validating the solution as being scalable, compliant with regulations, and suitable for the development of privacy-preserving enterprise AI.