Federated multimodal artificial intelligence framework for privacy-preserving smart infrastructure towards smart city

Le résumé fourni par la source

Smart infrastructure systems in urban areas, such as transportation systems, energy networks, and building control systems, integrate vast amounts of multimodal data that are essential for operational intelligence and for addressing the challenges of developing smart and sustainable cities. But centralized learning paradigms pose significant privacy risks, and current federated paradigms treat data modalities uniformly, failing to account for cross-modal heterogeneity and infrastructure-specific constraints. In this paper, a novel federated multimodal artificial intelligence framework to realize privacy-preserving smart infrastructure intelligence is introduced, dubbed FedSmartInfra. The proposed solution includes a heterogeneity-aware multimodal feature learning scheme exploiting graph transformer-based spatiotemporal representations and registered infrastructure clients based on the explicit representation of their aggregation history; a dynamic differential privacy-preserving aggregation protocol, which adapts the strength of the perturbation based on the contribution of the clients and their projection to non-IIDness; and a heterogeneity-aware learning scheme handling non-IID data distribution across infrastructure clients. The findings show that FedSmartInfra’s performance in anomaly detection accuracy is 94.2% in experimental testing for publicly available data sets provided on Kaggle and UCI, with the number of communications reduced by 34.2% compared to baseline approaches, and it can remain stable for different privacy budgets and poisoning attacks. The framework offers an extensible architecture for real-time processing and monitoring of different kinds of infrastructure sensing streams, which would facilitate secure, theoretically privacy-preserving, scalable, and interpretable intelligence in heterogeneous smart infrastructure environments.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.