Privacy-Preserving Federated Learning Framework for Sustainable Smart City Infrastructure Analytics

Abstract

The instrumentation of urban environments with dense Internet of Things (IoT) sensor networks has created a fundamentally new challenge for municipal data governance: how to harness the predictive value of geographically dispersed infrastructure data without consolidating sensitive operational information in centralised repositories that are simultaneously privacy-invasive and strategically vulnerable. This paper introduces PPFL-SCIA (Privacy-Preserving Federated Learning for Smart City Infrastructure Analytics), a governance-ready distributed learning framework that enables collaborative model training across heterogeneous urban infrastructure nodes whilst providing mathematically rigorous (ε, δ)-differential privacy guarantees. The framework integrates three interlocking privacy mechanisms — an adaptive sensitivity-classified Gaussian noise mechanism, a threshold-based secure gradient aggregation protocol and a lightweight lattice-based update verification scheme — within a bandwidth-efficient federated optimisation pipeline. A structured gradient compression module achieves a 64.8% reduction in per-round communication volume through dynamic sparsity scheduling and a resilience-aware client coordination protocol maintains stable convergence under simulated conditions of 35% node unavailability and severe cross-city data heterogeneity. Empirical evaluation across six UK urban deployments — London, Edinburgh, Cardiff, Sheffield, Nottingham and Leicester — over a 30-month operational window yielded anomaly detection accuracy of 93.1%, energy demand forecast mean absolute percentage error of 4.3% and flood-risk infrastructure alerting F1 of 0.907, with all tasks satisfying (ε ≤ 1.0, δ ≤ 10⁻⁵) privacy budgets. Relative to centralised and non-private federated baselines, PPFL-SCIA reduces infrastructure operational expenditure by 22.9%, unplanned downtime by 28.4% and average model convergence time by 38.6%.