Communication-Efficient Split Federated Learning for Edge LLM Training via Joint Cutting-Point and Resource Allocation
A communication-and-computation-efficient split federated learning framework for training large language models at the 6G network edge. It establishes theoretically that a smaller client-side model (cutting point) improves convergence, then formulates a joint Cutting-point selection, Communication and Computation resource allocation (CCC) problem minimizing convergence-rate-weighted latency under privacy constraints, solving the resulting mixed-integer nonlinear program by integrating a Double Deep Q-Network with convex optimization. Outperforms parallel split learning and traditional SFL.
Communication-and-Computation Efficient Split Federated Learning: Gradient Aggregation and Resource
Split Federated Learning (SFL) trains a neural network across many resource-limited clients and a server by splitting the model at a cutting-point layer: clients compute the lower layers and transmit…