A Comprehensive Survey of Incentive Mechanism for Federated Learning

Rongfei Zeng, Chao Zeng, Xingwei Wang, Bo Li, Xiaowen ChuarXiv 2021 · OpenReview

Abstract: Federated learning utilizes various resources provided by participants to collaboratively train a global model, which potentially address the data privacy issue of machine learning. In such promising paradigm, the performance will be deteriorated without sufficient training data and other resources in the learning process. Thus, it is quite crucial to inspire more participants to contribute their valuable resources with some payments for federated learning. In this paper, we present a comprehensive survey of incentive schemes for federate learning. Specifically, we identify the incentive problem in federated learning and then provide a taxonomy for various schemes. Subsequently, we summarize the existing incentive mechanisms in terms of the main techniques, such as Stackelberg game, auction, contract theory, Shapley value, reinforcement learning, blockchain. By reviewing and comparing some impressive results, we figure out three directions for the future study.

The Incentive Problem in Federated Learning

Federated Learning (FL) allows multiple parties to train a machine learning model collaboratively without sharing their raw data. While this solves major privacy hurdles, it introduces a "free-rider" or participation problem. Training models consumes significant local resources: battery, CPU cycles, bandwidth, and the inherent cost of data privacy. Without a clear reward, rational participants have little reason to join or provide high-quality data. Furthermore, participants with poor data can actually degrade the global model.

The Federated Learning Incentive Cycle

As illustrated in the workflow above, the system involves a cycle of parameter downloads, local training, and updates. The challenge lies in designing a mechanism that accurately evaluates contributions, selects the best nodes, and allocates payments fairly to ensure the resulting model is superior to one trained in isolation.

Taxonomy of Incentive Mechanisms

The authors provide a systematic taxonomy, breaking down incentive mechanisms into three core sub-problems:

  1. Contribution Evaluation: Determining how much a participant's data actually helped (e.g., using Shapley Value).
  2. Node Selection: Choosing which subset of available participants to engage for a specific round.
  3. Payment Allocation: Deciding the form and amount of reward (cash, reputation, or model access).
Taxonomy of Incentive Mechanisms in FL

The survey summarizes the main techniques used to solve these problems, ranging from Game Theory (Stackelberg games, Auctions, and Contract Theory) to Reinforcement Learning and Blockchain for decentralized automation.

Key Insights and Methodological Analysis

Performance Improvement (PI) as a Metric

A central argument of the paper is that traditional economic properties like "Fairness" or "Budget Balance" are insufficient. In FL, an incentive mechanism is only successful if it leads to measurable Performance Improvement. If paying participants doesn't result in a more accurate or faster-converging model, the mechanism has failed.

The Computational Cost of Fairness

While the Shapley Value is the most mathematically fair way to evaluate contribution, its exponential complexity (O(N!)) makes it nearly impossible to use in large-scale mobile networks without significant approximation.

Information Asymmetry

Participants usually have private information about their costs and data quality. Mechanisms like Contract Theory are highlighted as effective ways for a server to screen participants by offering a menu of options that force them to reveal their true types.

Key Results

The paper synthesizes findings from various state-of-the-art schemes:

  • Efficiency Gains: Frameworks like FMore (multi-dimensional auction) were shown to speed up FL training by reducing rounds by 51.3% and improving accuracy by 28% on specific CNN tasks.
  • Robustness: The FAIR framework demonstrated that quality-aware incentives allow FL to outperform standard "greedy" selection methods even when up to 80% of the provided data is mislabeled.

Limitations and Future Work

The authors identify several gaps in the current research:

  • High Computational Overhead: Many current schemes (especially those involving RL or Blockchain) are too heavy for resource-constrained mobile devices. Future work needs lightweight incentive designs.
  • Neglect of Cross-Silo FL: Most research focuses on mobile users. However, the incentives for large organizations to share data are more complex and less explored.
  • Multi-Dimensional Metrics: Future mechanisms should consider a multi-dimensional approach, accounting for data distribution (Non-IIDness), communication latency, and hardware reliability simultaneously.

Summary updated

Abstract

Federated learning utilizes various resources provided by participants to collaboratively train a global model, which potentially address the data privacy issue of machine learning. In such promising paradigm, the performance will be deteriorated without sufficient training data and other resources in the learning process. Thus, it is quite crucial to inspire more participants to contribute their valuable resources with some payments for federated learning. In this paper, we present a comprehensive survey of incentive schemes for federate learning. Specifically, we identify the incentive problem in federated learning and then provide a taxonomy for various schemes. Subsequently, we summarize the existing incentive mechanisms in terms of the main techniques, such as Stackelberg game, auction, contract theory, Shapley value, reinforcement learning, blockchain. By reviewing and comparing some impressive results, we figure out three directions for the future study.
BibTeX
@misc{zeng2021comprehensive,
  author = {Rongfei Zeng and Chao Zeng and Xingwei Wang and Bo Li and Xiaowen Chu},
  title = {{A Comprehensive Survey of Incentive Mechanism for Federated Learning}},
  year = {2021},
  eprint = {2106.15406},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2106.15406},
}
Figure