Towards Realistic Mechanisms That Incentivize Federated Participation and Contribution

TL;DR

This work tackles the lack of participation incentives and the free-rider problem in cross-device federated learning by designing RealFM, a mechanism that models device utility with a non-linear payoff $\phi_i(a)$ and data costs, while providing model-accuracy rewards $a^r$ and monetary rewards $R$. It extends the FL formulation to heterogeneous, non-sharing data and introduces an accuracy-shaping function $\gamma_i(m)$ to incentivize data contributions beyond local optima, funded by a dynamic marginal reward $r(\bm{m})$ and a server-retained share $p_m$. The authors prove feasibility, IR, and the existence of a unique Nash equilibrium under their mechanism, effectively eliminating free-riding and increasing data contributions. Empirical results on CIFAR-10 and MNIST show real-world performance gains, with server and device utilities improving by orders of magnitude compared to baselines across varied data distributions and costs, underscoring the practicality of realistic incentive designs in FL.

Abstract

Edge device participation in federating learning (FL) is typically studied through the lens of device-server communication (e.g., device dropout) and assumes an undying desire from edge devices to participate in FL. As a result, current FL frameworks are flawed when implemented in realistic settings, with many encountering the free-rider dilemma. In a step to push FL towards realistic settings, we propose RealFM: the first federated mechanism that (1) realistically models device utility, (2) incentivizes data contribution and device participation, (3) provably removes the free-rider dilemma, and (4) relaxes assumptions on data homogeneity and data sharing. Compared to previous FL mechanisms, RealFM allows for a non-linear relationship between model accuracy and utility, which improves the utility gained by the server and participating devices. On real-world data, RealFM improves device and server utility, as well as data contribution, by over 3 and 4 magnitudes respectively compared to baselines.

Figures (12)

• Figure 1: Federated Mechanism Diagram.(A) Decision Phase for Device Participation. Devices decide whether they want to participate in the mechanism. If so, the quantity of data points used by each agent $m_i$ is sent to the server (no data is shared). We note that rational devices would participate in RealFM; the utility gained by participating is never less than what agents attain locally. (B) Federated Training Phase. Devices upload their updates and receive feedback from the server in an iterative manner. (C) Accuracy & Monetary Reward Distribution Phase. Upon completion of federated training, the server distributes both accuracy $a^r(m_i)$ and monetary $R(m_i)$ rewards to device $i$. These rewards, the crux of RealFM, incentivize device participation and data contribution.
• Figure 2: Local Optimal Data Contribution for Varying Payoff Functions. We compare optimal data contribution across different payoff functions. Realistic power payoff functions, $\phi_i(a) = \frac{1}{(1-a)^2} - 1$, result in greater optimal contribution compared to linear payoff functions, $\phi_i(a) = a$. We define $\hat{a}_i(m)$ as in Equation \ref{['eq:gen-bound']}, with $a_{opt}^i = 0.95$ and multiple $k$ values.
• Figure 3: Improved Server Utility on CIFAR-10 & MNIST.RealFM increases server utility on CIFAR-10 (top row) and MNIST (bottom row) for $16$ devices compared to baselines. RealFM achieves upwards of 5 magnitudes more utility than a FL version of karimireddy2022mechanisms, denoted as LinearRealFM, across both uniform and various heterogeneous Dirichlet data distributions (left: uniform, center: D-0.6, right: D-0.3) as well as non-uniform costs (C) and accuracy payoff functions (P).
• Figure 4: Increased Federated Contribution on CIFAR-10 & MNIST.RealFM incentivizes devices to use more local data during federated training on CIFAR-10 (top row) and MNIST (bottom row) for $16$ devices compared to relevant baselines. RealFM achieves upwards of 4 magnitudes more federated contributions than LinearRealFM across both uniform and various heterogeneous Dirichlet data distributions (left: uniform, center: D-0.6, right: D-0.3) as well as non-uniform costs (C) and accuracy payoff functions (P).
• Figure 5: Utility Functions for Varying Cost and Payoff Functions. Using both linear, $\phi_i(a) = a$, and power, $\phi_i(a) = \frac{1}{(1-a)^2} - 1$, payoff functions, we compare how device utilities change with rising costs. Once marginal costs $c_i$ become too high, the utility is always negative and devices will not collect data for training. We use $\hat{a}_i(m)$ as defined in Equation \ref{['eq:gen-bound']}, with $a_{opt}^i = 0.95$ and $k=1$.

Theorems & Definitions (21)

• definition 1: Feasible Mechanism
• definition 2: Individual Rationality (IR)
• theorem 1: Existence of Pure Equilibrium
• remark 1
• remark 2: Attainability
• remark 3: Server Accuracy
• remark 4: Heterogeneous Distributions
• theorem 2: Optimal Local Data Collection
• theorem 3: Accuracy Shaping Guarantees
• remark 5