Federated Learning and Differential Privacy for Cross-Firm Construction-Materials Supply-Chain Disruption Forecasting: A Quantitative Analysis of Privacy-Utility Trade-Offs

Abstract

Construction firms forecast materials disruptions from their own order histories, because delivery records reveal suppliers, volumes and vulnerabilities to competitors, yet no single firm's short history can estimate the regional component of disruption risk. Federated learning with differential privacy promises a shared forecaster without shared records, but its costs for consortia with short, correlated order panels, rare outcomes and a common regional shock are unmeasured. This study quantifies the trade-offs in a controlled simulation with a fully specified generator of order-line panels: federations of 5 to 100 firms, 6 to 24 periods of history, a shared regional factor and three levels of heterogeneity trained logistic regression and neural network forecasters of next-period disruption under local-only, centralised, federated, sample-level differentially private (ε 0.5 to 8) and client-level differentially private (ε 4 to 64) regimes across 1,840 runs, with Rényi-DP accounting, out-of-time evaluation (AUC, calibration, watch-list capture, expected cost) in a base and a shock scenario, and a membership-inference attack. Federation recovered nearly all of the pooling benefit (0.3 AUC points below centralised training at moderate heterogeneity) and most helped small firms and short histories; under strong heterogeneity, local fine-tuning of the shared forecaster beat both shared and local models. Sample-level privacy cost the logistic forecaster 1.7 points at ε = 4 and 4.2 at ε = 1 with ten periods of history, less with longer ones, and crippled the network; client-level privacy at conventional budgets was unattainable with ten firms (19.6 points at ε = 4). A regional shock at the horizon left ranking almost intact but cut decision value through calibration (FedAvg saving 33% to 24%; local 26% to 13%); the federated forecaster kept more of its value than local ones, and local recalibration did not help. The attack gained nothing measurable against the shared logistic forecasters but did against isolated local models. The findings are translated into design guidance, a governance workflow, a decision tree and a roadmap. Limitations are synthetic data, small models and a single black-box attack.