Heterogeneity-Aware Federated Aggregation for Privacy-Preserving Urban Incident Detection: Convergence Analysis and Multi-Benchmark Evaluation

Urban safety systems must process heterogeneous sensor streams in real time while respecting citizen privacy, yet mainstream deep learning centralises raw data for training, which conflicts with regulations such as India's Digital Personal Data Protection Act and the GDPR, and models the city as a flat grid that ignores road topology and incident propagation. We present Heterogeneity-Aware Federated Aggregation (HAFA), a federated learning method for non-IID, temporally correlated urban data, paired with a Dynamic Spatio-Temporal Graph Neural Network (DST-GNN). HAFA derives Bayesian uncertainty estimates from each client's local posterior through Monte Carlo dropout and uses them as adaptive aggregation weights, giving greater influence to lower-variance updates; it carries a convergence guarantee under bounded heterogeneity and reduces to FedAvg when uncertainties are equal. Across five public benchmarks HAFA reaches an incident-detection F1 of 89.3 percent, only 2.4 points below a centralised bound, converges in 47 communication rounds against 89 for FedAvg, and cuts gradient transmission by 54.6 percent, with dynamic graph topology adding a statistically significant gain (t = 4.73, p < 0.001). HAFA thus offers a deployable, privacy-safe foundation for next-generation smart-city safety.