APEW-Fed: adaptive privacy-aware ensemble weighting for federated anomaly detection in enterprise resource planning systems.

Enterprise Resource Planning (ERP) systems are high-value targets for fraud owing to the sensitive financial and operational data they process. Anomaly detection in such systems faces a fundamental tension: centralising data for accurate detection conflicts with privacy regulations such as the European Union's General Data Protection Regulation (GDPR) and the United States' Sarbanes-Oxley Act (SOX). We present APEW-Fed, a federated ensemble framework that resolves this tension through three provably complementary mechanisms. First, Adaptive Privacy-Aware Ensemble Weighting (APEW) frames model fusion as a privacy-penalised optimisation problem and derives closed-form dynamic weights that up-weight privacy-efficient, high-confidence detectors. Second, Feature-Sensitivity Calibrated Differential Privacy (FS-CDP) partitions features into sensitivity tiers and allocates per-tier Gaussian noise budgets that provably minimise total noise variance subject to a global Rényi DP constraint. Third, Federated Anomaly Score Calibration (FASC) reconciles heterogeneous score distributions across clients via differentially private quantile sketches under secure aggregation. Integrating Isolation Forest, a federated autoencoder trained with per-step DP-SGD, adaptive DBSCAN, and gradient boosting, we evaluate on 243,531 real-world ERP transactions with 5000 expert-annotated anomalies and validate on two public benchmarks: the IEEE-CIS Fraud Detection dataset and the NSL-KDD network intrusion dataset. APEW-Fed achieves 93.5% F1-score at [Formula: see text] RDP-only 2.7 percentage points below the non-private centralised baseline-while reducing communication cost by 94% versus FedAvg. On a strictly unbiased uniform-only evaluation subset (2500 transactions, no score-based enrichment), APEW-Fed achieves 92.1% F1, confirming that the headline result is not an artefact of the annotation sampling design. Membership inference attack success is 51.2%, near-random guessing, confirming strong empirical privacy.