Blockchain-Aided Wireless Federated Learning: Resource Allocation and Client Scheduling

Jun Li, Weiwei Zhang, Kang Wei, Guangji Chen, Feng Shu, Wen Chen, Shi Jin

TL;DR

This paper introduces DRC-BDFL, a novel blockchain-aided federated learning (FL) framework for wireless networks that tackles the challenges of trust, single points of failure, and resource limitations. By jointly optimizing resource allocation and client scheduling using Lyapunov optimization, it significantly reduces training latency while maintaining comparable accuracy to existing methods.

What Problem Does This Paper Address?

Traditional centralized federated learning (FL) systems, while effective for distributed model training, suffer from a critical trust issue and a single point of failure due to their reliance on a central server. To address this, blockchain-aided decentralized FL (BDFL) has emerged, leveraging blockchain's inherent security and decentralization. However, deploying BDFL in resource-constrained wireless environments (like IoT devices) presents new challenges: limited bandwidth, processing power, and energy. This paper aims to solve these issues by formulating a dynamic stochastic optimization problem to minimize average training delay, considering constraints on energy budgets and client participation.

Methods

The core of this work is the DRC-BDFL (Dynamic Resource Allocation and Client Scheduling BDFL) algorithm. The authors tackle the complex, long-term mixed-integer nonlinear programming problem by employing the Lyapunov optimization framework. This framework transforms the problem into a per-communication-round deterministic optimization problem by introducing virtual queues to manage long-term constraints. The problem is then decoupled into three sub-problems: optimizing client computation frequency ($f_i(t)$), mining frequency ($f_{bloc}(t)$), and client scheduling ($S(t)$). These sub-problems are solved iteratively using Lagrange multipliers and a greedy approach based on latency. The algorithm aims to minimize training delay while respecting energy constraints and ensuring client participation rates.

Datasets

The experiments were conducted using two standard image datasets:

  • SVHN (Street View House Numbers): A dataset of house numbers from Google Street View images, used for digit recognition.
  • CIFAR-10: A dataset of 60,000 RGB images across 10 categories, commonly used for image classification.

The paper also notes that experiments were performed under a Non-IID (Non-Independent and Identically Distributed) scenario, achieved by dividing the datasets according to a Dirichlet distribution. This simulates more realistic heterogeneous data distributions across clients.

Key Insights and Ideas

  • Blockchain for Decentralized FL in Wireless: The paper demonstrates a practical application of BDFL in wireless settings, highlighting its potential to overcome the limitations of centralized FL.
  • Lyapunov Optimization for Complex Constraints: The successful application of Lyapunov optimization to manage long-term fairness and stability constraints in a dynamic, resource-constrained BDFL environment is a key methodological contribution.
  • Joint Optimization is Crucial: The results emphasize that simultaneously optimizing computation, communication, and client scheduling is essential for minimizing training latency in BDFL.
  • Trade-offs in BDFL: The work implicitly highlights the trade-offs between training latency, energy consumption, and model accuracy, showing that DRC-BDFL effectively balances these factors.

Key Results

  • Reduced Training Delay: DRC-BDFL significantly reduces training latency compared to baseline algorithms (Random, Round Robin, Channel State-based scheduling). Specifically, it achieved reductions of 9.24% on SVHN and 12.47% on CIFAR-10.
  • Comparable Accuracy: The proposed algorithm achieves test accuracy comparable to the baseline methods.
  • Energy Constraint Satisfaction: Experimental results show that the algorithm effectively operates within the specified maximum energy consumption constraints ($E_{max}$).
  • Theoretical Convergence: The paper provides a theoretical analysis of the algorithm's convergence, deriving an upper bound of $O(1/T)$ for the convergence rate.

Limitations and Future Work

  • Future Work - Enhanced Privacy: The authors suggest exploring stronger privacy-preserving techniques, such as homomorphic encryption and secure multi-party computation, to further enhance data security and client privacy in BDFL.
  • Future Work - 6G Integration: Integrating FL with 6G technology to combine sensing and communication functions into a unified framework is proposed as another promising research direction. This could lead to cross-layer resource management for integrated sensing and communications (ISAC).
  • Future Work - Multi-Object Recognition: The potential of FL to address multi-object multi-task recognition problems in ISAC systems is also mentioned.
  • Implicit Limitation - Complexity: While the paper proposes an efficient algorithm, the underlying problem is inherently complex due to its stochastic and combinatorial nature. The iterative optimization process, though effective, might still have computational overheads in very large-scale deployments.

Summary updated

Abstract

Federated learning (FL) based on the centralized design faces both challenges regarding the trust issue and a single point of failure. To alleviate these issues, blockchain-aided decentralized FL (BDFL) introduces the decentralized network architecture into the FL training process, which can effectively overcome the defects of centralized architecture. However, deploying BDFL in wireless networks usually encounters challenges, such as limited bandwidth, computing power, and energy consumption. Driven by these considerations, a dynamic stochastic optimization problem is formulated to minimize the average training delay by jointly optimizing the resource allocation and client selection under the constraints of limited energy budget and client participation. We solve the long-term mixed integer nonlinear programming problem by employing the tool of Lyapunov optimization and thereby propose the dynamic resource allocation and client scheduling BDFL (DRC-BDFL) algorithm. Furthermore, we analyse the learning performance of DRC-BDFL and derive an upper bound for convergence regarding the global loss function. Extensive experiments conducted on the SVHN and CIFAR-10 data sets demonstrate that the DRC-BDFL achieves comparable accuracy to the baseline algorithms while significantly reducing the training delay by 9.24% and 12.47%, respectively.
BibTeX
@article{li2024blockchain,
  author = {Jun Li and Weiwei Zhang and Kang Wei and Guangji Chen and Feng Shu and Wen Chen and Shi Jin},
  title = {{Blockchain-Aided Wireless Federated Learning: Resource Allocation and Client Scheduling}},
  journal = {IEEE Internet of Things Journal},
  year = {2024},
  url = {https://openreview.net/forum?id=9fdCEIcTKb},
}
Figure