Incentive Mechanism Design for Joint Resource Allocation in Blockchain-based Federated Learning

Zhilin Wang, Qin Hu, Ruinian Li, Minghui Xu, Zehui Xiong

Introduction

Federated Learning (FL) has emerged as a powerful paradigm for training AI models while preserving data privacy. By allowing edge devices to train models locally, FL avoids the risks associated with centralized data storage. However, traditional FL architectures often rely on a central server, which introduces a single point of failure and potential trust issues. Integrating blockchain technology into FL (BCFL) addresses these concerns by providing a decentralized, immutable ledger for model updates and incentive distribution.

Despite its benefits, BCFL introduces a significant challenge: resource competition. Mobile edge devices, acting as both trainers and miners, must allocate their limited CPU cycles between these two computationally intensive tasks. This paper proposes a sophisticated incentive mechanism to optimize this joint resource allocation.

System Architecture and Problem Formulation

The proposed system involves two primary entities: Model Owners (MOs) and Clients. The MO initiates the learning task and provides rewards, while Clients contribute their local data for training and participate in the blockchain consensus process (mining).

System architecture of the blockchain-based federated learning framework, illustrating the interaction between the Model Owner and the decentralized network.

The core problem is a "balancing act." If a client devotes too much power to training, they may fail to complete mining within the required timeframe. Conversely, excessive mining power reduces the quality or speed of model training. Furthermore, the MO often operates under "incomplete information," not knowing the exact data volume or computational capacity of each client. Without a precise incentive mechanism, clients might under-contribute or report false information to maximize their own profit.

The Stackelberg Game Approach

To solve this, the authors model the interaction as a two-stage Stackelberg game. In the first stage, the MO (the leader) sets the unit prices for training and mining rewards. In the second stage, the Clients (the followers) determine their optimal CPU cycle allocation based on these prices to maximize their own utility.

Key Innovations

  • Fairness via Shapley Value: The mechanism uses the Shapley Value to bound rewards, ensuring that payments are proportional to a client's actual contribution to the global model's accuracy, preventing a "rich-get-richer" scenario.
  • Truthfulness under Incomplete Information: By incorporating Incentive Compatibility (IC) and Individual Rationality (IR) constraints, the mechanism ensures that clients maximize their utility only by reporting their data volume truthfully.
  • Optimal Resource Allocation: Using backward induction and Karush-Kuhn-Tucker (KKT) conditions, the authors derive a Nash Equilibrium where both the MO and clients achieve their best possible outcomes.

Experimental Results and Insights

Numerical simulations with 50 clients demonstrate the effectiveness of the proposed mechanism. The results indicate that the optimal strategies derived from the Stackelberg game significantly outperform random allocation methods.

Comparison of total utility for clients and the Model Owner under different strategy combinations, showing the superiority of the optimal-optimal approach.

As shown in the figures, the "Optimal vs. Optimal" strategy (where both the MO and clients follow the proposed mechanism) leads to a linear increase in utility as the network scales. This highlights the scalability of the algorithm, which maintains a time complexity of $O(n)$, making it suitable for large-scale IoT networks. Furthermore, the results show that increasing the unit price for training successfully stimulates clients to provide more computing power, directly improving model performance.

Conclusion

This research provides a robust framework for managing the dual-task demands of BCFL. By aligning the incentives of Model Owners and edge devices through game theory and fairness metrics, the system ensures sustainable and efficient decentralized learning. Future work will look into dynamic network environments where clients join or leave mid-round, and the impact of communication overhead on the equilibrium.

Summary updated

BibTeX
@article{wang2023incentive,
  author = {Zhilin Wang and Qin Hu and Ruinian Li and Minghui Xu and Zehui Xiong},
  title = {{Incentive Mechanism Design for Joint Resource Allocation in Blockchain-based Federated Learning}},
  journal = {IEEE Trans. Parallel Distributed Syst.},
  year = {2023},
  doi = {10.1109/tpds.2023.3253604},
}
Figure