Differential Privacy and Blockchain-Empowered Decentralized Graph Federated Learning-Enabled UAVs for Disaster Response

Natural disasters, such as earthquakes, can cause damage to critical infrastructures and limit access to vital information, making it difficult for disaster response teams to respond effectively. Unmanned aerial vehicles (UAVs) have the potential to aid and provide real-time information for disaster response teams, however, the need to process distributed learning for huge amounts of interconnected nodes in a graph network poses several challenges. First, distributed learning in graph networks for UAVs is still an open issue, making it difficult to train and share models on such networks. Second, such a network can leak privacy-sensitive information, making it harder to ensure data security. To address these challenges, we propose, in this article, a novel privacy and blockchain-empowered UAVs-enabled decentralized graph federated learning (DPBE-DGFL) framework for disaster response. The framework includes three phases: 1) local model training utilizing stochastic gradient descent with differential privacy; 2) model weights integrity authentication using blockchain to ensure secure and efficient sharing of model weights; and 3) final validator selection and model weights aggregation using a Dedicated Proof-of-Stake (DPoS), consensus mechanism to ensure efficient and decentralized consensus while maintaining security and integrity. Our DPBE-DGFL framework was evaluated using extensive simulations on EMNIST and real-world disaster data sets from Tonga. The results show that it offers a promising solution for privacy-preserving federated learning in graph networks, balancing privacy protection and model accuracy while maintaining latency, communication, and computational efficiency.