QI-DPFL: Quality-Aware and Incentive-Boosted Federated Learning with Differential Privacy

TL;DR

QI-DPFL tackles privacy leakage and participation incentives in federated learning with heterogeneous data. It integrates an Earth Mover's Distance-based quality-aware client selection with a rho-zero-concentrated differential privacy framework and a two-stage Stackelberg incentive game to coordinate server-client interactions. The authors derive a Stackelberg Nash Equilibrium, providing closed-form solutions for optimal client privacy budgets and server rewards, and prove equilibrium existence. Experiments on MNIST, CIFAR-10, and EMNIST demonstrate competitive accuracy with lower costs, especially under non-IID conditions, highlighting the method's practical impact for privacy-preserving, incentive-aware FL.

Abstract

Federated Learning (FL) has increasingly been recognized as an innovative and secure distributed model training paradigm, aiming to coordinate multiple edge clients to collaboratively train a shared model without uploading their private datasets. The challenge of encouraging mobile edge devices to participate zealously in FL model training procedures, while mitigating the privacy leakage risks during wireless transmission, remains comparatively unexplored so far. In this paper, we propose a novel approach, named QI-DPFL (Quality-Aware and Incentive-Boosted Federated Learning with Differential Privacy), to address the aforementioned intractable issue. To select clients with high-quality datasets, we first propose a quality-aware client selection mechanism based on the Earth Mover's Distance (EMD) metric. Furthermore, to attract high-quality data contributors, we design an incentive-boosted mechanism that constructs the interactions between the central server and the selected clients as a two-stage Stackelberg game, where the central server designs the time-dependent reward to minimize its cost by considering the trade-off between accuracy loss and total reward allocated, and each selected client decides the privacy budget to maximize its utility. The Nash Equilibrium of the Stackelberg game is derived to find the optimal solution in each global iteration. The extensive experimental results on different real-world datasets demonstrate the effectiveness of our proposed FL framework, by realizing the goal of privacy protection and incentive compatibility.

Figures (6)

• Figure 1: The framework of QI-DPFL consists of the following procedures: Step 1: The central server initializes the global model and publishes the task request. Step 2: The clients submit their data distributions to the central server and are selected by EMD metric. Step 3: Each selected client collects private database $\mathcal{D}_{i}$ and performs local training. Step 4: Once finishing the local training, each client uploads the perturbed model parameter $\nabla \widetilde{F}_{i}(\boldsymbol{w_{i}(t)})$ to the central server. Step 5: The central server offers a reward $\mathcal{R}_{t}$ to all participating clients to compensate the cost of each selected participant suffered. Step 6: Upon aggregating local parameters, the global model updates and is further verified on the testing set. The training process concludes when the required accuracy or a preset number of iterations is reached. Otherwise, the central server redistributes the updated global model to clients for the next iteration (Steps 3-6).
• Figure 2: The relationship between the privacy budget $\rho$ and model accuracy on MNIST and EMNIST dataset
• Figure 3: The performance of different models on EMNIST dataset with IID data distribution.
• Figure 4: The performance of different models on EMNIST dataset with Non-IID data distribution.
• Figure 5: The performance of different models on Cifar10 dataset with IID data distribution.

Theorems & Definitions (8)

• Definition 1
• Corollary 1
• Theorem 1
• Theorem 2
• Definition 2
• Theorem 3
• Lemma 1
• Remark 1