PaperlayerView exact revision on arXiv

Source-linked AI summary

A Comprehensive Survey of Incentive Mechanism for Federated Learning

Rongfei Zeng, Chao Zeng, Xingwei Wang, Bo Li, Xiaowen Chu

arXiv:2106.15406v1cs.LGcs.GT

TL;DR

Federated learning relies on participants’ resources, yet insufficient contributions can impair training, making incentives important. This survey organizes incentive mechanisms by their techniques and functions, and concludes that performance improvement, application-specific constraints, and cross-silo FL deserve attention.

  • Problem

    Federated learning can lose training performance when participants do not provide sufficient data, bandwidth, or computation, motivating incentives for their resources.

    Source: §I · p. 1

  • Method

    The survey formulates the incentive problem, organizes mechanisms by their components and techniques, and compares studies while identifying their assumptions.

    Source: §I · pp. 1–2

  • Results

    The survey identifies FL performance improvement as an incentive-design priority and highlights constraints in MEC, 5G/B5G, and IoT settings and limited study of cross-silo FL.

    Source: §I · p. 2

  • Takeaways & Limitations

    Future incentive designs should account for application-specific constraints and give more attention to cross-silo FL.

    Source: §I · p. 2

Abstract

from arXiv · show

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.

I. INTRODUCTION

Federated learning keeps raw data local while participants send model parameters for aggregation, but training can suffer when participants lack sufficient resources. This survey frames incentives around improving FL performance and organizes existing mechanisms and future research priorities.

I. INTRODUCTION

  • Insufficient training data, bandwidth, or computation can deteriorate model accuracy and training speed, motivating payments for participants’ resources.Payments include cash, business reputation, or a share of the well-trained model.Source: §I · p. 1
  • This survey complements an earlier overview by offering a different taxonomy and summary perspective alongside its comparison of existing incentive mechanisms.The earlier survey groups studies by client contribution, reputation, and resource allocation.Source: §I · p. 1
  • Incentive design should improve FL training performance, since a few high-quality participants can outperform many low-contributing participants.The authors identify contribution evaluation, node selection, and payment allocation as components of incentive mechanisms.Source: §I · p. 1
  • The survey maps techniques to functions: Shapley value commonly evaluates contribution, while Stackelberg games, auctions, and contract theory often allocate payments.Reinforcement learning and blockchain also support node selection, contribution evaluation, and robustness; techniques can serve multiple functions, and assumptions matter in mechanism design.Source: §I · p. 2
  • The survey identifies three future priorities: link incentives to FL performance, address constraints in MEC, 5G/B5G, and IoT scenarios, and study cross-silo FL.The authors report that current studies neglect cross-silo FL.Source: §I · p. 2

II. FEDERATED LEARNING, PROBLEM STATEMENT, AND THE TAXONOMY

Federated learning trains a global model through participant updates, while incentive mechanisms formalize how participants contribute, are selected, and are paid. The section argues that these mechanisms should improve FL performance and describes their component structure and desired properties.

A. Federated Learning

  • Participants train locally and upload model parameters for global aggregation, which is redistributed for another round until accuracy or a time threshold is reached.Participants may be devices, edge nodes, or IoT devices in cross-device FL, or companies in cross-silo FL.Source: A. Federated Learning · p. 2

B. Problem Statement of Incentive Mechanisms in FL

  • The formal mechanism chooses contributions and desirable properties or objectives; participants earn payment minus cost, while the model owner’s utility is reduced by payments.A participant with zero contribution is not selected for training.Source: B. Problem Statement of Incentive… · p. 2
  • Nash equilibrium is presented as a desirable strategy because no player has an incentive to choose a different resource provision.The paper also notes dominant-strategy, Bayesian, and perfect Bayesian equilibrium as alternatives.Source: B. Problem Statement of Incentive… · pp. 2–3
  • The listed properties include incentive compatibility, individual rationality, and fairness: truthful reporting is optimal, participant profits are nonnegative, and predefined fairness functions are maximized.The paper identifies fairness as key to sustainable collaboration in FL.Source: B. Problem Statement of Incentive… · p. 3
  • Pareto efficiency maximizes social surplus; collusion resistance rules out subgroup gains from fraud, and budget balance limits total payments to the available budget.Source: B. Problem Statement of Incentive… · p. 3
  • FL incentive mechanisms add performance improvement as a requirement, spanning accuracy, training speed, communication overhead, and computational costs.The paper motivates this requirement by noting that high-quality participants can outperform many low-quality participants.Source: B. Problem Statement of Incentive… · p. 3
  • The mechanism comprises interdependent contribution evaluation, node selection, and payment allocation; a single technique can address multiple components.Contribution evaluation measures participant contributions, node selection chooses qualified participants, and payment allocation sets their payments.Source: B. Problem Statement of Incentive… · p. 3

C. The Taxonomy of Incentive Mechanisms

The taxonomy organizes incentive schemes along five dimensions, spanning application settings, FL phases, techniques, sub-problems, and information symmetry.

C. The Taxonomy of Incentive Mechanisms

  • The taxonomy classifies incentive schemes by application setting, FL phase, main technique, sub-problem, and information symmetry.Fig. 3 presents the proposed taxonomy.Source: C. The Taxonomy of Incentive Mechanisms · p. 3, Figure 3 · p. 4
  • Cross-device and cross-silo FL differ in scale and resources; most surveyed studies address cross-device settings, while fewer target cross-silo FL.Cross-device settings involve many resource-constrained nodes, whereas cross-silo settings involve fewer organizations with substantial data and distributed centers.Source: C. The Taxonomy of Incentive Mechanisms · pp. 3–4
  • Training and prediction are distinct FL phases with different goals and assumptions, leading to different incentive schemes.Source: C. The Taxonomy of Incentive Mechanisms · p. 4
  • Stackelberg games, auctions, and contracts mainly support node selection and payment allocation; Shapley value measures contribution, while blockchain and reinforcement learning serve as auxiliary techniques.Source: C. The Taxonomy of Incentive Mechanisms · p. 4
  • The sub-problems are contribution evaluation, node selection, and payment allocation, while schemes also differ by whether information is incomplete, weakly complete, or complete.Some studies address multiple sub-problems simultaneously.Source: C. The Taxonomy of Incentive Mechanisms · p. 4

III. INCENTIVE MECHANISM OF FL

This section presents a technical survey of state-of-the-art incentive mechanisms for federated learning. It organizes the discussion around the main techniques used in existing schemes.

III. INCENTIVE MECHANISM OF FL

  • The section summarizes state-of-the-art incentive mechanisms for FL from a technical perspective.Source: §III · p. 4
  • The survey separately presents schemes based on Shapley value, Stackelberg games, auctions, and contracts.Source: §III · p. 4
  • It also covers reinforcement learning, blockchain, and other techniques used in incentive mechanisms.Source: §III · p. 4

A. Shapley Value

The section explains Shapley value as a way to assess participant contributions through marginal utility, while reviewing its computational and privacy challenges and proposed approaches.

A. Shapley Value

  • Shapley value measures participant contribution through weighted average marginal contributions, addressing why data volume alone can misrepresent value [22],.The marginal contribution is the change in coalition utility when a participant joins.Source: A. Shapley Value · p. 4
  • The worked example computes coalition marginal contributions and averages each participant’s corresponding values to obtain Shapley values.Source: A. Shapley Value · p. 4
  • Shapley value has computational complexity O(N!), requiring an exponential number of utility calculations [34].Sampling, group-testing, probabilistic estimation, and information-sharing approaches have been studied to approximate it.Source: A. Shapley Value · pp. 4–5
  • Two gradient-based methods calculate enterprise participants’ model utilities by reconstructing models after training or aggregating round-wise contributions.Both trade additional storage of local gradients for avoiding model retraining.Source: A. Shapley Value · p. 5
  • FedCoin uses competing miners to approximate Shapley value, selects the result nearest the miners’ average, and records payment allocation in a blockchain [25].The approach bounds computational resources for reaching consensus.Source: A. Shapley Value · p. 5
  • Direct Shapley computation can expose private feature values or distributions; Shapley group value was proposed for a two-participant vertical FL case.The passage notes that computation becomes intractable as the number of participants increases.Source: A. Shapley Value · p. 5
  • Shapley value is also used for participant behavior analysis and feature selection, alongside contribution measurement,.Source: A. Shapley Value · p. 5

B. Stackelberg Game

The section describes Stackelberg games as sequential incentive models and surveys their role assignments, product definitions, information assumptions, and reported FL outcomes.

B. Stackelberg Game

  • In a Stackelberg game, a leader moves first and followers respond; backward induction can compute the Nash equilibrium [16],,.Source: B. Stackelberg Game · p. 5
  • Most FL studies make the model owner the sole leader and participants the followers, but one mobile scenario assigns multiple devices as leaders and the owner as follower.The review says the differing market assumptions remain unjustified.Source: B. Stackelberg Game · p. 5
  • Stackelberg incentives treat training data, computation, accuracy, or privacy as products exchanged for rewards or compensation.Examples include CPU provision under a leader’s budget and privacy compensation proportional to clients’ privacy budgets.Source: B. Stackelberg Game · p. 5
  • With incomplete or stochastic information, computing the Nash equilibrium can add N(N−1) incentive-compatibility constraints; a DRL scheme uses PPO to adjust strategies under ambiguous contribution measurement.The scheme entails high computation overhead for edge nodes and uses offline simulation training.Source: B. Stackelberg Game · pp. 5–6
  • For MDS-coded training, the optimal recovery threshold is reported to scale linearly with participant count, while a cooperative relay design significantly reduces communication congestion, [18].Source: B. Stackelberg Game · p. 6

C. Auction

Auction-based FL incentives let participants bid resources and prices while the server selects winners and allocates payments. The survey describes multidimensional auction designs and reports results tied to training performance and data quality.

C. Auction

  • Auctions proceed from bid announcement and collection to winner determination and payment, with the server acting as auctioneer and participants as bidders.The survey describes auction-based incentives as useful for pricing, task allocation, and node selection.Source: C. Auction · p. 6
  • Because winner selection and payment allocation can be computationally intractable, surveyed approaches include randomized greedy methods and neural-network approximations,, [10].The survey also describes DRLA using graph neural network embeddings of participant conflicts for service allocation and payment.Source: C. Auction · pp. 6–7
  • FL auction designs often combine multiple resource attributes, either scalarized by a scoring function or treated separately during winner search,,.Examples include data size, computation, bandwidth, data distribution, and wireless-channel requests.Source: C. Auction · p. 6
  • In incentive design, proposed scheme FMore speeds up FL training by reducing training rounds by 51.3% on average and improves accuracy by 28% for the tested CNN.FMore is described as inspiring high-quality nodes with sufficient resources to join FL training.Source: C. Auction · p. 6
  • FAIR’s model accuracy outperforms the knapsack greedy mechanism when mislabeled data ranges from 20% to 80%.FAIR combines local learning-quality estimation for node selection with aggregation that accounts for data size and estimated quality.Source: C. Auction · p. 7

D. Contract Theory

The surveyed contract mechanisms address information asymmetry and heterogeneous resources by mapping participant types or contributions to rewards. Designs range from type-based contracts to multidimensional provisions.

D. Contract Theory

  • Public procurement contracts offer participants a menu from which they select an option suited to their private type, using self-revelation under information asymmetry.The information-revelation performance depends on the granularity of the contract list.Source: D. Contract Theory · p. 7
  • Under strongly incomplete information, the base station publishes reward-and-computation contracts, and participants choose their true type to maximize private utility [14].The publisher’s optimization uses monotonicity, incentive-compatibility, and individual-rationality constraints; the study is extended with reputation and blockchain.Source: D. Contract Theory · p. 7
  • A multi-dimensional contract study considers complete, weakly incomplete, and strongly incomplete information, using a two-part uniform contract for the potentially non-convex strongly incomplete case.Its stated performance approaches the original optimization problem’s optimum in that scenario.Source: D. Contract Theory · p. 7
  • For heterogeneous UAV resources, a two-step contract converts sensing and computation costs into a single-dimensional design, then adds compensation for traversal and transmission costs [31].Another design uses data size and communication time as decision variables and simplifies the incentive problem to a single-dimensional contract.Source: D. Contract Theory · p. 7

E. Reinforcement Learning

The survey presents reinforcement learning as an approximation approach for incentive decisions made repeatedly in FL. It groups existing schemes by discrete versus continuous action spaces.

E. Reinforcement Learning

  • In FL incentive design, an agent can repeatedly select nodes or allocate payments, observe participant responses as rewards, and seek improved training performance.The survey motivates RL partly because many incentive optimization problems are NP-hard and large participant populations make exact solutions difficult.Source: E. Reinforcement Learning · p. 8
  • For discrete actions, double DQN selects a predefined number of devices using Q values to counterbalance data-distribution bias.DQN is also used in a randomized multidimensional auction that selects winners while considering channel conflicts and local data distribution, improving social welfare over traditional reverse multidimensional auction while guaranteeing truthfulness and IR.Source: E. Reinforcement Learning · p. 8
  • For continuous decisions, PPO is used to allocate payments in incomplete-information Stackelberg games and to jointly maximize system performance and minimize costs,.The latter study focuses on federated-learning systems in real-world settings.Source: E. Reinforcement Learning · p. 8

F. Others

Beyond the main incentive techniques, the survey covers blockchain-based mechanisms and other settings involving cross-silo learning, prediction, fairness, and truthful reporting.

F. Others

  • Blockchain is used to support secure or privacy-preserving FL incentives, including DeepChain’s confidentiality, computation auditability, and party incentives.Other surveyed designs use blockchain for auditable learning, vote-based rewards, or commitments and strategic incentives against copying outputs.Source: F. Others · p. 8
  • A cross-silo scheme models organizations’ resources as non-excludable public goods and maximizes social welfare while satisfying individual rationality and budget balance.Its distributed algorithm is designed to obtain Nash-equilibrium strategies under perfect information.Source: F. Others · p. 8
  • For prediction-phase incentives, a Bayesian-game framework rewards truthful predictions using a divergence-based Bayesian Truth Serum, alongside truth discovery and trusted execution environments.The passage reports individual rationality, budget balance, and performance improvement for the framework.Source: F. Others · p. 8
  • FLI dynamically shares a given budget to maximize collective utility while minimizing inequality among data owners, targeting three forms of fairness,.The passage reports near-optimal collective utility while limiting data-owner regret.Source: F. Others · pp. 8–9
  • Other surveyed mechanisms include truthful reporting through information elicitation and truthful mechanisms for interdependent values with type-dependent action spaces [28], [50].The cited work on information elicitation studies cases with and without verification.Source: F. Others · p. 9

IV. COMPARISONS AND FUTURE STUDIES

The comparisons identify performance goals and design trade-offs across incentive schemes. They also point to limits in computational cost, metric scope, and integration of incentive subproblems.

A. Comparisons and Analysis

  • All node-selection schemes target FL performance, and some recent incentive schemes, including and, also target performance alongside IR, IC, and fairness.Source: A. Comparisons and Analysis · p. 9
  • Many studies use computationally expensive techniques such as reinforcement learning, blockchain, and Shapley value, creating overhead unsuitable for mobile or resource-constrained participants.The passage identifies both computational cost and communication overhead.Source: A. Comparisons and Analysis · p. 9
  • Single-dimensional metrics may not fit FL, where both participant resources and model performance should be measured across multiple dimensions.Source: A. Comparisons and Analysis · p. 9
  • Most current mechanisms target a single incentive aspect, while few jointly address node selection, payment allocation, and contribution evaluation despite their correlation.The survey argues these subproblems should be considered together systemically.Source: A. Comparisons and Analysis · p. 9

B. Future Studies of Incentive in FL

The survey proposes four future directions for FL incentives: low-cost performance improvement, cross-silo schemes, comprehensive designs, and new technologies for new scenarios. It also presents a taxonomy of incentive mechanisms and compares existing studies.

B. Future Studies of Incentive in FL

  • Future schemes should improve FL performance at low cost by encouraging participation, since recruiting participants without improving performance is not the incentive mechanism's goal.The survey also calls for lightweight schemes because resource-constrained nodes may hesitate to perform expensive incentive computations.Source: B. Future Studies of Incentive in FL · p. 9
  • Cross-silo FL needs distinct incentive methods because companies' and organizations' decision behavior differs from that of end users and mobile devices.The survey notes that FL applications are widespread in cross-silo settings, making incentives more critical there.Source: B. Future Studies of Incentive in FL · pp. 9–10
  • Future schemes should use multidimensional metrics and combine multiple goals and functions rather than focus on a single dimension.Source: B. Future Studies of Incentive in FL · p. 10
  • Graph neural networks, generative adversarial networks, and multi-agent reinforcement learning may apply to incentive design in scenarios such as mobile edge computing and 5G/B5G.Source: B. Future Studies of Incentive in FL · p. 10

V. CONCLUSION

  • The survey formulates the FL incentive problem, presents a taxonomy and design goals, summarizes mechanisms by technique, compares studies, and identifies four future directions.Source: §V · p. 10
Loading 2106.15406v1…