[Paper Review] Low-Latency Federated Learning over Wireless Channels with Differential Privacy

Kang Wei, Jun Li+6 more|arXiv (Cornell University)|Jun 20, 2021

Privacy-Preserving Technologies in Data42 references5 citations

TL;DR

This paper proposes a low-latency federated learning framework over wireless channels that jointly optimizes training delay, model performance, and client-level differential privacy using a multi-agent multi-armed bandit (MAMAB) approach. By leveraging Lyapunov drift for constraint handling and solving a max-min bipartite matching problem via a modified Hungarian algorithm and a greedy matching with better alternative (GMBA), the method achieves low-complexity, near-optimal performance with a regret bound growing logarithmically with communication rounds.

ABSTRACT

In federated learning (FL), model training is distributed over clients and local models are aggregated by a central server. The performance of uploaded models in such situations can vary widely due to imbalanced data distributions, potential demands on privacy protections, and quality of transmissions. In this paper, we aim to minimize FL training delay over wireless channels, constrained by overall training performance as well as each client's differential privacy (DP) requirement. We solve this problem in the framework of multi-agent multi-armed bandit (MAMAB) to deal with the situation where there are multiple clients confornting different unknown transmission environments, e.g., channel fading and interferences. Specifically, we first transform the long-term constraints on both training performance and each client's DP into a virtual queue based on the Lyapunov drift technique. Then, we convert the MAMAB to a max-min bipartite matching problem at each communication round, by estimating rewards with the upper confidence bound (UCB) approach. More importantly, we propose two efficient solutions to this matching problem, i.e., modified Hungarian algorithm and greedy matching with a better alternative (GMBA), in which the first one can achieve the optimal solution with a high complexity while the second one approaches a better trade-off by enabling a verified low-complexity with little performance loss. In addition, we develop an upper bound on the expected regret of this MAMAB based FL framework, which shows a linear growth over the logarithm of communication rounds, justifying its theoretical feasibility. Extensive experimental results are conducted to validate the effectiveness of our proposed algorithms, and the impacts of various parameters on the FL performance over wireless edge networks are also discussed.

Motivation & Objective

• Address the challenge of minimizing training delay in wireless federated learning under constrained model performance and individual client differential privacy (DP) requirements.
• Tackle the inherent uncertainty in wireless environments—such as channel fading and interference—where client transmission conditions are unknown and time-varying.
• Design a scalable, distributed learning framework that maintains privacy and efficiency in dynamic, real-world wireless edge networks.
• Balance trade-offs between communication latency, model convergence, and individual privacy guarantees in a multi-client wireless FL system.
• Develop a theoretical regret bound to validate the long-term performance and feasibility of the proposed MAMAB-based learning algorithm.

Proposed method

• Transform long-term constraints on model performance and client DP into virtual queues using the Lyapunov drift technique to enable dynamic optimization.
• Model the client scheduling and resource allocation problem as a multi-agent multi-armed bandit (MAMAB) framework to handle unknown and non-stationary wireless channel conditions.
• Estimate rewards for each client-action pair using the Upper Confidence Bound (UCB) method to balance exploration and exploitation in the MAMAB setting.
• Formulate the MAMAB problem at each communication round as a max-min bipartite matching problem to maximize the minimum reward across clients.
• Propose two efficient solutions: a modified Hungarian algorithm for optimal matching with high complexity, and a greedy matching with better alternative (GMBA) for low-complexity with minimal performance loss.
• Derive a theoretical upper bound on the expected regret of the MAMAB framework, showing linear growth with log of communication rounds, confirming its asymptotic optimality.

Experimental results

Research questions

• RQ1 How can training delay in wireless federated learning be minimized while satisfying both global model performance and individual client differential privacy constraints?
• RQ2 What is the optimal client scheduling and resource allocation strategy in a dynamic wireless environment with unknown channel states and interference?
• RQ3 How can the trade-off between computational complexity and performance be balanced in large-scale, privacy-preserving FL systems?
• RQ4 What theoretical performance guarantees can be provided for a MAMAB-based FL framework under time-varying wireless conditions and privacy constraints?
• RQ5 How do key system parameters—such as client data distribution, DP budget, and channel quality—affect the convergence and latency of the FL process?

Key findings

• The proposed GMBA algorithm achieves near-optimal performance with significantly reduced computational complexity compared to the exact Hungarian method, making it suitable for real-time deployment.
• The theoretical regret bound grows logarithmically with the number of communication rounds, indicating that the algorithm converges efficiently and maintains long-term performance stability.
• The framework successfully maintains client-level differential privacy while minimizing training delay, even under imbalanced data distributions and varying channel conditions.
• Extensive simulations validate that the proposed method outperforms baseline scheduling policies (e.g., random, round-robin) in terms of convergence speed and final model accuracy.
• The impact of DP parameters and channel quality on FL performance is quantitatively analyzed, showing that higher DP budgets improve model accuracy but increase communication load.
• The Lyapunov drift-based constraint handling effectively maintains both training performance and DP compliance across all clients over time.