Abstract
The rapid proliferation of smart city infrastructures has led to The fast growth of infrastructures of smart cities has brought about an increase in the amount and diversity of data produced by the Internet of Things (IoT). Intelligent transportation systems, healthcare devices, environmental monitoring sensors, and utility networks have all contributed to this increase. However, despite the ability to conduct analytics through the federated learning (FL) approach and the graph neural network (GNN), the currently available techniques are vulnerable to inference attacks, graph structure leakage, large communication cost, and poor performance in resource-limited edge nodes. To overcome these drawbacks, we present Privacy-Preserving Adaptive Federated Hypergraph Neural Network (PAFHGNN) for privacy-preserving IoT data analytics in 6G smart cities. Adaptive hypergraph representation learning and hierarchical federated optimization have been considered in the proposed framework for learning high-order interactions in a heterogeneous collection of IoT devices while maintaining decentralized learning process. Differential privacy-based gradient perturbation, secure aggregation, and blockchain-powered model verification techniques have been included in the design for securing node data as well as graph topology from malicious attacks. Moreover, a communication-efficient aggregation scheme is designed for minimizing communication overhead among heterogeneous edge devices without degrading the performance of the machine learning models. Extensive experimentation on an IoT dataset of a smart city reveals that the proposed PAFHGNN attains 98.9% accuracy, 98.3% precision, 98.1% recall, and 98.2% F1-score with a communication overhead of 291 MB. The framework outperforms the existing federated graph learning framework in terms of performance improvement of up to 6.8% as well as communication overhead reduction of 20.3%.
Keywords: Federated Learning, Hypergraph Neural Network, Smart City IoT, Privacy Preservation, 6G Networks