IMFLKD: an incentive mechanism for decentralized federated learning based on knowledge distillation

Knowledge Distillation-based Federated Learning (KD-FL) has garnered significant attention as one of the core technical pathways for next-generation Federated Learning (FL), owing to its communication efficiency, privacy preservation, and strong robustness. Meanwhile, to further reduce reliance on a central server, blockchain-enabled KD-FL architectures have become a research hotspot. However, designing an effective incentive mechanism that encourages participants to consistently contribute high-quality knowledge remains a fundamental challenge for ensuring the system's long-term sustainability. To address this issue, this paper proposes an Incentive Mechanism for decentralized FL based on Knowledge Distillation (IMFLKD). First, we design a two-stage evaluation method, combining smart contract-based label aggregation and peer-wise comparison, that enables accurate client model quality estimation and fair reward allocation without increasing time complexity. Second, we establish a multi-dimensional dynamic reputation system based on the Subjective Logic model, incorporating metrics such as data quality, activity level, and stability to identify high-value participants and incentivize sustained, high-quality contributions across multiple FL rounds rather than short-term opportunistic behavior. Finally, we integrate these components into a decentralized, blockchain-enabled KD-FL framework. Experimental results demonstrate that IMFLKD achieves superior performance in contribution assessment accuracy, computational overhead, and resilience against malicious attacks, showcasing strong practicality and reliability.