Split Federated Learning Architectures for High-Accuracy and Low-Delay Model Training

Yiannis Papageorgiou; Yannis Thomas; Ramin Khalili; Iordanis Koutsopoulos

cs.LGMar 9, 2026

Split Federated Learning Architectures for High-Accuracy and Low-Delay Model Training

Authors: Yiannis Papageorgiou, Yannis Thomas, Ramin Khalili, Iordanis Koutsopoulos

Organizations: Department of Informatics, Athens University of Economics and Business, Athens, Greece · Huawei Technologies Duesseldorf Gmbh, Munich, Germany

Abstract

Can we find a network architecture for ML model training so as to optimize training loss (and thus, accuracy) in Split Federated Learning (SFL)? And can this architecture also reduce training delay and communication overhead? While accuracy is not influenced by how we split the model in ordinary, state-of-the-art SFL, in this work we answer the questions above in the affirmative. Recent Hierarchical SFL (HSFL) architectures adopt a three-tier training structure consisting of clients, (local) aggregators, and a central server. In this architecture, the model is partitioned at two partitioning layers into three sub-models, which are executed across the three tiers. Despite their merits, HSFL architectures overlook the impact of the partitioning layers and client-to-aggregator assignments on accuracy, delay, and overhead. This work explicitly captures the impact of the partitioning layers and client-to-aggregator assignments on accuracy, delay and overhead by formulating a joint optimization problem. We prove that the problem is NP-hard and propose the first accuracy-aware heuristic algorithm that explicitly accounts for model accuracy, while remaining delay-efficient. Simulation results on public datasets show that our approach can improve accuracy by 3%, while reducing delay by 20% and overhead by 50%, compared to state-of-the-art SFL and HSFL schemes.

Figures & tables

Fig. 1: Example network topology that presents a network architecture with three clients, where each client trains the model in collaboration with the server. Client 1, being computationally stronger, is selected as a local aggregator. The model is partitioned into three sub-models (a, b, and c) as defined by the aggregator layer and the cut layer. The clients, local aggregators and server train the first (a), middle (b) and last (c) sub-model respectively.

Fig. 2: Test accuracy versus training epochs for different cut layers during the training of AlexNet and VGG-11 models across 100 clients with local-loss learning.

Reference	Straggler effect				Backward locking effect	Accuracy-awareness
Reference	Topology-awareness	Helpers	Cut layer selection	Aggregation strategy	Local-loss	Joint optimization
[ 10 ] , [ 16 ] , [ 15 ]	✓		✓
[ 14 ] , [ 17 ]	✓		✓	✓
[ 6 ] , [ 24 ] , [ 25 ]					✓
[ 18 ] , [ 19 ]		✓	✓
[ 8 ] , [ 9 ]		✓	✓

TABLE I: Comparison of existing SFL schemes and the proposed accuracy-aware solution.

Fig. 3: Batch processing with one server and 3 clients, where client 3 acts as local aggregator. The figure depicts the training tasks (rectangles) executed by each client in time (x-axis). The horizontal blue lines mark the execution zone of each node; tasks within this zone are executed by the corresponding node; multi-threaded execution is implied when multiple tasks in the same zone overlap in time. For each task, we present the trained sub-model, the training “direction” (FP or BP) and the trained layers, e.g.,

W^{w}_{3}

FP(1,h)

encodes that FP of layers 1 to h in the weak-side sub-model of client 3. The arrows symbolize the network transmission of activations and gradients in the training chain.

$a_{l}$	Weights of layer $l$ (bits)
$B$	Size of data batch
$Q$	Number of batch processing executions within a epoch
$g_{l}$	Activations/gradients of layer $l$ (bits)
$E$	Number of epochs
$T_{i}$	Training delay of phase $i,i\in[1,2,3]$
$f_{l}$	Computation workload of layer $l$ (FLOPS)

TABLE II: Notation Table

Fig. 4: Test accuracy versus (a) training delay and (b) communication overhead achieved by AA HSFL-ll and baseline schemes during the training of the AlexNet model on the MNIST dataset.

Fig. 5: Test accuracy versus (a) training delay and (b) communication overhead achieved by AA HSFL-ll and baseline schemes during the training of the VGG-11 model on the CIFAR-10 dataset.

Fig. 6: Test accuracy versus (a) training delay and (b) communication overhead achieved by AA HSFL-ll and baseline schemes during the training of the VGG-19 model on the CIFAR-10 dataset.

Fig. 7: Test accuracy versus (a) training delay and (b) communication overhead achieved by AA HSFL-ll and baseline schemes during the training of the ResNet-101 model on the CINIC-10 dataset.

Model	Acc. Target(%)	Delay Gain (%)	Overhead Gain (%)
AlexNet	85	7.25	57.14
AlexNet	90	22.00	55.56
AlexNet	94	20.00	50.00
VGG-11	65	35.00	37.50
VGG-11	70	28.41	30.00
VGG-11	75	11.34	34.62

TABLE III: Relative gains (%) achieved by AA-HSFL-ll in terms of training delay and communication overhead for various accuracy targets compared to the best performing baseline scheme.

Model	$h$	$v$	$\lambda$	Avg. clients
AlexNet	$3$	$5$	$0.2$	5.5
VGG-11	$3$	$8$	$0.2$	4
VGG-19	$4$	$15$	$0.25$	3
ResNet-101	$8$	$24$	$0.3$	2.25

TABLE IV: Selected aggregator layer (

h

), cut layer (

v

) fraction of clients operating as local aggregators (

\lambda

), and average clients assigned per aggregator during the training of the four models using the proposed algorithm.

Fig. 8: Test accuracy obtained by AA HSFL-ll during training the AlexNet model for various numbers of local aggregators

\lambda N

Fig. 9: Test accuracy versus communication overhead obtained by AA HSFL-ll during training the ResNet-101 model for various numbers of local aggregators

\lambda N

$\lambda$	0.10	0.25	0.30	0.50	0.80
AlexNet	(3, 5)	(2, 5)	-	(2, 5)	(2, 5)
ResNet-101	(9, 24)	-	(8, 24)	(9, 26)	(12, 26)

TABLE V: Aggregator layer and cut layer (

h

v

) selection during the training of the AlexNet and ResNet-101 models for various fractions of clients operating as local aggregators

\lambda

Model	$\gamma$	Accuracy (% )
VGG-11	$2$	76
VGG-11	$7.5$	76.8
VGG-11	$15$	75.2
ResNet-101	$2$	73
ResNet-101	$7.5$	73.2
ResNet-101	$15$	73

TABLE VI: Accuracy achieved by our approach for various client heterogeneity degrees

\gamma

during the training of the VGG-11 and ResNet-101 models.

Model	$\gamma$	$h$	$v$	$\lambda$
VGG-11	$2$	$3$	$8$	$0.10$
VGG-11	$7.5$	$4$	$8$	$0.20$
VGG-11	$15$	$5$	$8$	$0.30$
ResNet-101	$2$	$11$	$26$	$0.15$
ResNet-101	$7.5$	$8$	$24$	$0.30$
ResNet-101	$15$	$6$	$24$	$0.30$

TABLE VII: Aggregator layer

h

and fractions of clients operating as local aggregators

\lambda

selection during the training of the VGG-11 and ResNet-101 models for various client heterogeneity degrees

\gamma

Model	$N$	sub-optimality (%)	speedup ( $x$ Times)
VGG-19	30	0	10.5
VGG-19	60	0	24
VGG-19	100	10.5	36
ResNet-101	30	0	10.5
ResNet-101	60	0	22
ResNet-101	100	12	40

TABLE VIII: Sub-optimality and speedup achieved by our algorithm when compared to the exhaustive search for various number of clients

N

during the training of VGG-19 and ResNet-101 models.

Type of change	Model	Fixed sol. (%)	New sol. (%)
Extra 15 % tasks	VGG-19	+5	+2
Extra 30 % tasks	VGG-19	+12	+5
Reduced TX rate	VGG-19	+18	+7
Extra 15 % tasks	ResNet-101	+14	+9
Extra 30 % tasks	ResNet-101	+24	+11
Reduced TX rate	ResNet-101	+7	+3

TABLE IX: Comparison of fixed and proposed solutions under different change types during the training of VGG-19 and ResNet-101 models.

Split Federated Learning

Split federated learning divides models between clients and servers to train collaboratively while keeping raw data local.

See all 37 papers

ASFL: An Adaptive Model Splitting and Resource Allocation Framework for Split Federated Learning
Feb 19, 2026Chuiyang Meng, Ming Tang, Vincent W. S. WongSplit Federated Learning Energy-Efficient ML
Department of Electrical and Computer Engineering, The University of British Columbia, Vancouver, Canada · Department of Computer Science and Engineering, Southern University of Science and Technology, China
Efficient Resource Optimization for Split Federated Learning
Aug 18, 2026Wei Wei, Xianhao ChenSplit Federated Learning
Department of Electrical and Computer Engineering, the University of Hong Kong, Pok Fu Lam, Hong Kong SAR, 999077
HybridFLow: SDN-Orchestrated Client Partitioning for Hybrid Federated Learning
Sep 9, 2026Osama Abu Hamdan, Rabin Pandey, Hao Che +2Asynchronous Federated Learning Federated Learning
University of Texas at Arlington · Missouri University of Science and Technology · Meta Platforms, Inc.

Feb 19, 2026J/K move · Enter open · S save

ASFL: An Adaptive Model Splitting and Resource Allocation Framework for Split Federated Learning

Chuiyang Meng, Ming Tang, Vincent W. S. Wong

Department of Electrical and Computer Engineering, The University of British Columbia, Vancouver, Canada · Department of Computer Science and Engineering, Southern University of Science and Technology, China

Split Federated Learning Energy-Efficient ML Wireless Resource Allocation

Open paper AskPDF arXiv

Federated learning (FL) enables multiple clients to collaboratively train a machine learning model without sharing their raw data. However, the limited computation resources of the clients may result in a high delay and energy consumption on training. In this paper, we propose an adaptive split federated learning (ASFL) framework over wireless networks. ASFL exploits the computation resources of the central server to train part of the model and enables adaptive model splitting as well as resource allocation during training. To optimize the learning performance (i.e., convergence rate) and efficiency (i.e., delay and energy consumption) of ASFL, we theoretically analyze the convergence rate and formulate a joint learning performance and resource allocation optimization problem. Solving this problem is challenging due to the long-term delay and energy consumption constraints as well as the coupling of the model splitting and resource allocation decisions. We propose an online optimization enhanced block coordinate descent (OOE-BCD) algorithm to solve the problem iteratively. Experimental results show that when compared with five baseline schemes, our proposed ASFL framework converges faster and reduces the total delay and energy consumption by up to 75% and 80%, respectively.

Abstract

Split Federated Learning

Explore similar work

ASFL: An Adaptive Model Splitting and Resource Allocation Framework for Split Federated Learning

Efficient Resource Optimization for Split Federated Learning

HybridFLow: SDN-Orchestrated Client Partitioning for Hybrid Federated Learning