UAVs and Blockchain Synergy: Enabling Secure Reputation-based Federated Learning in Smart Cities

Unmanned aerial vehicles (UAVs) can be used as drones’ edge Intelligence to assist with data collection, training models, and communication over wireless networks. UAV use for smart cities is rapidly growing in various industries, including tracking and surveillance, military defense, managing healthcare delivery, wireless communications, and more. In traditional machine learning techniques, an enormous amount of sensor data from UAVs must be shared to central storage to perform model training, which poses serious privacy risks and risks of misuse of information. The federated learning technique (FL), which can be applied to UAVs, is a promising means of collaboratively training a global model while retaining local access to sensitive raw data. Despite this, FL is a significant communication burden for battery-constrained UAVs due to local model training and global synchronization frequency. In this article, we address the major challenges associated with UAV-based FL for smart cities, including single-point failure, privacy leakage, scalability, and global model verification. To tackle these challenges, we present a differentially private federated learning framework based on Accumulative Reputation-based Selection (ARS) for the edge-aided UAV network that utilizes blockchains to prevent single-point failures where we switched from central control to decentralized control, Interplanetary File System (IPFS) for off-chain model storage and their respective hash-keys on-chain to ensure model integrity. Due to IPFS, the size of the blockchain will be reduced, and local differential privacy will be applied to prevent privacy leakages. In the proposed framework, an aggregator will be selected based on its ARS score and model verification by the validators. After most validators approve it, it will be available for use. Several parameters are taken into consideration during evaluation, including accuracy, precision, recall, F1-score, and time consumption. It also evaluates the number of edge computers vs test accuracy, the number of edge computers vs time consumption for global model convergence, and the number of rounds vs test accuracy. This is done by considering two benchmark datasets: MNIST and CIFAR-10. The results show that the proposed work preserves privacy while achieving high accuracy. Moreover, it is scalable to accommodate many participants.