Incentive mechanism design in blockchain-based federated learning over edge clouds

Federated learning (FL) deployed over edge computing environments is a promising paradigm for preserving data privacy while enabling real-time, decentralized artificial intelligence. However, conventional FL faces critical challenges, including severe communication bottlenecks, vulnerability to single points of failure, and a lack of transparency in model aggregation. While integrating blockchain technology into FL systems mitigates these security risks and ensures traceability, the resulting architecture demands significant computational and communication resources from participating edge servers and client devices. Because these participants are rational and resource-constrained, the success of blockchain-enabled FL relies heavily on robust incentive mechanisms that can motivate contributions to both model training and the blockchain consensus process. This dissertation investigates and designs comprehensive, game-theoretic incentive mechanisms for blockchain-enabled federated edge learning systems, progressively addressing challenges from architectural latency to information asymmetry. First, this research addresses the latency and straggler issues inherent in synchronous FL by proposing an incentive mechanism for a semi-asynchronous blockchain-based federated edge learning system. The resource pricing interaction between edge servers and task publishers is modeled as a two-stage Stackelberg game. To achieve optimal strategies efficiently, an iterative algorithm based on the Alternating Direction Method of Multipliers (ADMM) is developed, ensuring convergence to a unique Nash equilibrium. This foundation is then extended to a more complex, three-tier cloud-edge-client Hierarchical Federated Learning (HFL) architecture. Here, a three-stage Stackelberg game is constructed to model the resource pricing among clients, edge servers, and cloud publishers, which is solved using a hybrid ADMM and backward induction algorithm. Both approaches demonstrate that optimizing under complete information significantly improves system utility and model accuracy compared to baseline strategies. Finally, the dissertation tackles the realistic constraint of incomplete information alongside the integration of coded computing to further mitigate straggler effects. A novel hybrid incentive framework is proposed for blockchain-enabled coded HFL. At the lower layer, multidimensional contract theory is utilized to design personalized contracts that elicit truthful participation from clients with heterogeneous data volumes, privacy sensitivities, and computational capacities. At the upper layer, the reward allocation between the task publisher and edge servers is formulated as a single-leader multi-follower Stackelberg game, solved via a decentralized reinforcement learning (RL) algorithm that operates without requiring participants to reveal private data. Extensive theoretical analysis and simulations confirm that this hybrid approach guarantees incentive compatibility, individual rationality, and stable equilibrium even under severe information asymmetry. Ultimately, this dissertation establishes a rigorous theoretical algorithm design framework for building secure, scalable, and self-sustaining decentralized AI systems.