Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Incentivize Contribution and Learn Parameters Too: Federated Learning with Strategic Data Owners

Drashthi Doshi, Aditya Vema Reddy Kesari, Avishek Ghosh, Swaprava Nath, Suhas S Kowshik

TL;DR

This work tackles incentive design in federated learning by treating data contributors as strategic agents who decide how much data to sample. It introduces Updated Parameter Best Response Dynamics (UPBReD), which jointly updates contributions and the global model, and proves convergence to a pure Nash equilibrium. To achieve truthfulness, it adds ex-post incentive-compatible payments via T-UPBReD, and to reach the socially optimal outcome it proposes 2P-UPBReD, a two-phase, budget-balanced mechanism that drives full data contribution and attains $w^{OPT}$. Empirical results on CIFAR-10, FEMNIST, and Twitter show 2P-UPBReD closely matches FedAvg performance while overcoming strategic manipulation, with robustness enhancements such as trimmed-mean aggregation to resist adversarial behavior. The framework provides a practical path to incentive-aware FL with provable convergence and welfare guarantees, applicable to real-world, heterogeneous data settings.

Abstract

Classical federated learning (FL) assumes that the clients have a limited amount of noisy data with which they voluntarily participate and contribute towards learning a global, more accurate model in a principled manner. The learning happens in a distributed fashion without sharing the data with the center. However, these methods do not consider the incentive of an agent for participating and contributing to the process, given that data collection and running a distributed algorithm is costly for the clients. The question of rationality of contribution has been asked recently in the literature and some results exist that consider this problem. This paper addresses the question of simultaneous parameter learning and incentivizing contribution in a truthful manner, which distinguishes it from the extant literature. Our first mechanism incentivizes each client to contribute to the FL process at a Nash equilibrium and simultaneously learn the model parameters. We also ensure that agents are incentivized to truthfully reveal information in the intermediate stages of the algorithm. However, this equilibrium outcome can be away from the optimal, where clients contribute with their full data and the algorithm learns the optimal parameters. We propose a second mechanism that enables the full data contribution along with optimal parameter learning. Large scale experiments with real (federated) datasets (CIFAR-10, FEMNIST, and Twitter) show that these algorithms converge quite fast in practice, yield good welfare guarantees and better model performance for all agents.

Incentivize Contribution and Learn Parameters Too: Federated Learning with Strategic Data Owners

TL;DR

This work tackles incentive design in federated learning by treating data contributors as strategic agents who decide how much data to sample. It introduces Updated Parameter Best Response Dynamics (UPBReD), which jointly updates contributions and the global model, and proves convergence to a pure Nash equilibrium. To achieve truthfulness, it adds ex-post incentive-compatible payments via T-UPBReD, and to reach the socially optimal outcome it proposes 2P-UPBReD, a two-phase, budget-balanced mechanism that drives full data contribution and attains . Empirical results on CIFAR-10, FEMNIST, and Twitter show 2P-UPBReD closely matches FedAvg performance while overcoming strategic manipulation, with robustness enhancements such as trimmed-mean aggregation to resist adversarial behavior. The framework provides a practical path to incentive-aware FL with provable convergence and welfare guarantees, applicable to real-world, heterogeneous data settings.

Abstract

Classical federated learning (FL) assumes that the clients have a limited amount of noisy data with which they voluntarily participate and contribute towards learning a global, more accurate model in a principled manner. The learning happens in a distributed fashion without sharing the data with the center. However, these methods do not consider the incentive of an agent for participating and contributing to the process, given that data collection and running a distributed algorithm is costly for the clients. The question of rationality of contribution has been asked recently in the literature and some results exist that consider this problem. This paper addresses the question of simultaneous parameter learning and incentivizing contribution in a truthful manner, which distinguishes it from the extant literature. Our first mechanism incentivizes each client to contribute to the FL process at a Nash equilibrium and simultaneously learn the model parameters. We also ensure that agents are incentivized to truthfully reveal information in the intermediate stages of the algorithm. However, this equilibrium outcome can be away from the optimal, where clients contribute with their full data and the algorithm learns the optimal parameters. We propose a second mechanism that enables the full data contribution along with optimal parameter learning. Large scale experiments with real (federated) datasets (CIFAR-10, FEMNIST, and Twitter) show that these algorithms converge quite fast in practice, yield good welfare guarantees and better model performance for all agents.
Paper Structure (12 sections, 5 theorems, 44 equations, 4 figures, 5 algorithms)

This paper contains 12 sections, 5 theorems, 44 equations, 4 figures, 5 algorithms.

Table of Contents

  1. Introduction
  2. Our contributions
  3. Related work
  4. Preliminaries
  5. Learning with arbitrary effective costs
  6. Truthful elicitation with payment rules
  7. Application of \ref{['thm:convergence-genl']}: effective costs are only personal
  8. Welfare maximization at Nash equilibrium
  9. Convergence to the NE $(w^{\textnormal{OPT}},s^{\max})$
  10. Adversarial agents
  11. Experiments
  12. Conclusions and future work

Key Result

Theorem 1

Under assump:conv-NE, UPBReD (algo:main_algo) converges to a Nash equilibrium. Formally, for a given $\epsilon > 0$, for every initial value $(w^0,s^0)$ of algo:main_algo, the gradients $\|g(w^T,s^T,\mu^T)\|<\epsilon$ and $\|\tilde{g}(w^T,s^T)\|<\epsilon$, for all $T \geqslant T_0(w^0,s^0)$, when th

Figures (4)

  • Figure 1: Shaded regions show the feasible choices of $\gamma$ for $n=2,L=1$ and $\tilde{P}$ and $\lambda$ as shown. The dashed lines show the boundary of the regions in the legends. A similar set of choices is true for $\eta$.
  • Figure 2: The three rows in this figure correspond to the experiments with the datasets CIFAR-10, FEMNIST, and Twitter respectively. The y-axis in every plot is the social welfare. The x-axis corresponds to the number of agents in the first plot, the choice of $\beta$ in the second, and the time taken (in seconds) in the third plot respectively in every row.
  • Figure 3: Utilities and payments received by agents for 2P-UPBReD based on their data contributions for FEMNIST (averaged over 10 runs).
  • Figure 4: Ratio of social welfare with adversaries and without adversaries.

Theorems & Definitions (15)

  • Definition 1: Nash equilibrium
  • Definition 2: Budget balance
  • Definition 3: Ex-Post Incentive Compatible (EPIC)
  • Theorem 1: UPBReD convergence to NE
  • proof
  • Theorem 2
  • proof
  • Lemma 1: Increasing utility
  • proof
  • Definition 4: $M$-smoothness
  • ...and 5 more