Communication-and-Computation Efficient Split Federated Learning: Gradient Aggregation and Resource Management
Yipeng Liang, Qimei Chen, Guangxu Zhu, Muhammad Kaleem Awan, Hao Jiang
TL;DR
This work addresses the communication and computation bottlenecks of Split Federated Learning (SFL) for edge devices by introducing SFL-GA, which enables dynamic server/client model splitting and aggregated gradients to reduce uplink/downlink traffic. A joint communication-and-computation co-design (CCC) combines a DDQN-based cutting-point policy with convex optimization for resource allocation, aiming to minimize convergence latency and privacy leakage. Theoretical convergence bounds reveal that smaller client-side models improve convergence at the cost of privacy risk, guiding the CCC formulation. Extensive simulations across MNIST, Fashion-MNIST, and CIFAR-10 show that SFL-GA achieves comparable accuracy with substantially lower communication overhead and latency than traditional SFL, PSL, and FL baselines, validating the practical impact for edge-LM training and inference.
Abstract
With the prevalence of Large Learning Models (LLM), Split Federated Learning (SFL), which divides a learning model into server-side and client-side models, has emerged as an appealing technology to deal with the heavy computational burden for network edge clients. However, existing SFL frameworks would frequently upload smashed data and download gradients between the server and each client, leading to severe communication overheads. To address this issue, this work proposes a novel communication-and-computation efficient SFL framework, which allows dynamic model splitting (server- and client-side model cutting point selection) and broadcasting of aggregated smashed data gradients. We theoretically analyze the impact of the cutting point selection on the convergence rate of the proposed framework, revealing that model splitting with a smaller client-side model size leads to a better convergence performance and vise versa. Based on the above insights, we formulate an optimization problem to minimize the model convergence rate and latency under the consideration of data privacy via a joint Cutting point selection, Communication and Computation resource allocation (CCC) strategy. To deal with the proposed mixed integer nonlinear programming optimization problem, we develop an algorithm by integrating the Double Deep Q-learning Network (DDQN) with convex optimization methods. Extensive experiments validate our theoretical analyses across various datasets, and the numerical results demonstrate the effectiveness and superiority of the proposed communication-efficient SFL compared with existing schemes, including parallel split learning and traditional SFL mechanisms.