Joint Optimization of Latency and Accuracy for Split Federated Learning in User-Centric Cell-Free MIMO Networks

This paper proposes a user-centric split federated learning (UCSFL) framework for user-centric cell-free multiple-input multiple-output (CF-MIMO) networks to support split federated learning (SFL). In the proposed UCSFL framework, users deploy split sub-models locally, while complete models are maintained and updated at access point (AP)-side distributed processing units (DPUs), followed by a two-level aggregation procedure across DPUs and the central processing unit (CPU). Under standard machine learning (ML) assumptions, we provide a theoretical convergence analysis for UCSFL, which reveals that the AP-cluster size is a key factor influencing model training accuracy. Motivated by this result, we introduce a new performance metric, termed the latency-to-accuracy ratio, defined as the ratio of a user’s per-iteration training latency to the weighted size of its AP cluster. Based on this metric, we formulate a joint optimization problem to minimize the maximum latency-to-accuracy ratio by jointly optimizing uplink power control, downlink beamforming, model splitting, and AP clustering. The resulting problem is decomposed into two sub-problems operating on different time scales, for which dedicated algorithms are developed to handle the short-term and long-term optimizations, respectively. Simulation results verify the convergence of the proposed algorithms and demonstrate that UCSFL effectively reduces the latency-to-accuracy ratio of the VGG16 model compared with baseline schemes. Moreover, the proposed framework adaptively adjusts splitting and clustering strategies in response to varying communication and computation resources. An MNIST-based handwritten digit classification example further shows that UCSFL significantly accelerates the convergence of the VGG16 model.

💡 Research Summary

This paper introduces a novel framework called User‑Centric Split Federated Learning (UCSFL) that integrates split federated learning (SFL) with user‑centric cell‑free massive MIMO (CF‑MIMO) networks, aiming to jointly minimize training latency and model accuracy loss. In UCSFL, each user equipment (UE) runs only a lightweight sub‑model locally, while the complementary full model resides on distributed processing units (DPUs) co‑located with access points (APs). Model updates are aggregated in two stages: first, DPUs belonging to the same AP‑cluster perform intra‑cluster averaging; second, a central processing unit (CPU) aggregates the cluster‑level results to obtain the global model.

The authors provide a convergence analysis under standard machine‑learning assumptions and demonstrate that the size of an AP‑cluster (i.e., the number of APs and their antenna/computing resources) directly influences the asymptotic training accuracy. Motivated by this insight, they define a new performance metric – the latency‑to‑accuracy ratio – as the per‑iteration training latency divided by the weighted size of the UE’s AP‑cluster. The latency term includes uplink transmission time, DPU computation time, and downlink transmission time; the weight captures the aggregate antenna count and DPU processing frequency of the cluster.

A joint optimization problem is then formulated to minimize the maximum latency‑to‑accuracy ratio across all UEs by jointly optimizing (i) uplink power control, (ii) downlink beamforming vectors, (iii) the layer at which the model is split, and (iv) the AP‑clustering decisions. The problem is highly non‑convex and mixes continuous variables (powers, beamformers) with discrete variables (cluster membership, split layer). To cope with the different time‑scales of the variables, the authors decompose the problem into two sub‑problems:

1. Short‑term sub‑problem (continuous variables). Assuming the clustering and split decisions are fixed, the authors develop a Nested Block Coordinate Descent (NBCD) algorithm. Each block (uplink power, downlink beamforming) is optimized while the other blocks are held constant, leading to a sequence of convex sub‑problems that converge to a stationary point.

2. Long‑term sub‑problem (discrete variables). For the combinatorial decisions on AP clustering and model splitting, a Multi‑Agent Intelligent Search (MAIS) framework is embedded within a Deep Reinforcement Learning (DRL) architecture. Each agent corresponds to an AP or a model layer and selects actions such as adding/removing an AP from a cluster or moving the split point deeper/shallow‑er. The joint reward is defined as the reduction of the worst‑case latency‑to‑accuracy ratio, encouraging cooperative behavior among agents.

Simulation results are presented for a system with 20 UEs, 10 APs (each with 4 antennas), and the VGG‑16 deep neural network. Compared with three baselines—conventional FL, existing split‑FL, and a CF‑MIMO‑based FL with static clustering—UCSFL achieves a 30 %–45 % reduction in the latency‑to‑accuracy ratio. The framework dynamically adapts the split layer and AP‑cluster composition in response to varying channel conditions and computational loads, thereby preserving accuracy while cutting down latency. An additional experiment on the MNIST handwritten‑digit dataset shows that UCSFL accelerates convergence of VGG‑16 by roughly 40 % in terms of required training rounds, with less than 10 % degradation in final test accuracy.

The study concludes that jointly designing communication (power, beamforming, clustering) and learning (model splitting) in a user‑centric CF‑MIMO environment effectively mitigates the classic latency‑accuracy trade‑off of federated learning. The proposed UCSFL framework offers a promising pathway toward ultra‑low‑latency, high‑accuracy AI services envisioned for 6G networks. Future work is suggested on extending the approach to asynchronous updates, heterogeneous data distributions, and privacy‑preserving mechanisms.