Latent Information Sharing for Accelerating Federated Learning

Seungjun Lee; Ensieh Khazaei; Dimitrios Hatzinakos; Baturalp Buyukates; Sunwoo Lee

cs.LGOct 1, 2026

Latent Information Sharing for Accelerating Federated Learning

Authors: Seungjun Lee, Ensieh Khazaei, Dimitrios Hatzinakos, Baturalp Buyukates, Sunwoo Lee

Organizations: Inha University · University of Toronto · University of Birmingham

Abstract

Federated learning (FL) is a communication-efficient distributed learning paradigm. However, client drift remains one of the most critical challenges, hindering the efficient training of a global model. In this study, we propose a novel latent information sharing scheme that directly mitigates data heterogeneity across clients. Our theoretical and empirical results show that sharing a small amount of hidden-layer activations significantly improves training efficiency while preserving convergence guarantees and data privacy. Furthermore, we compare our method with existing FL approaches designed to address client drift, including FedProx, SCAFFOLD, FedPVR, FedProto, and SplitFed, and demonstrate superior model accuracy under a fixed round budget without incurring excessive communication overhead. Overall, this work presents a promising new knowledge aggregation scheme and provides a comprehensive analysis of the impact of activation sharing on federated optimization.

Figures & tables

Figure 1: A schematic illustration of latent information sharing. Figure (a) shows the Latent Information Sharing step. The activation outputs from $\mathbf{y}$ , $I_{A}$ and $I_{B}$ , are exchanged between the two clients. Figure (b) shows the Tail Round in which the two clients run backpropagation only within $\mathbf{z}$ using both $I_{A}$ and $I_{B}$ . Figure (c) shows the random client pairing-based activation sharing.

Figure 2: Layer selection results. The green curve represents the cumulative variance of the input-side sub-model y, computed using Eq. 6 , while the red curve represents the system efficiency computed using Eq. 5 . The red horizontal dashed line indicates the midpoint of the efficiency range. Across various benchmarks, the intersection of the two curves indicates a balanced trade-off between system efficiency and variance reduction.

Method	CIFAR-10		CIFAR-100		SVHN		FEMNIST
Method	Acc. (%) $\uparrow$	Comm (MB) $\downarrow$	Acc. $\uparrow$	Comm $\downarrow$	Acc. $\uparrow$	Comm $\downarrow$	Acc. $\uparrow$	Comm $\downarrow$
FedAvg	$67.70\pm 0.33$	42.63	$48.31\pm 0.68$	9.20	$87.13\pm 0.16$	1.04	$83.54\pm 0.19$	1.54
FedProx	$67.91\pm 0.53$	42.63	$47.61\pm 0.27$	9.20	$88.27\pm 0.90$	1.04	$83.23\pm 0.29$	1.54
FedProto	$69.57\pm 1.58$	42.65	$48.08\pm 0.87$	9.69	$87.99\pm 0.15$	1.04	$84.11\pm 0.23$	1.60
MOON	$69.35\pm 0.83$	42.63	$46.49\pm 0.21$	9.20	$86.10\pm 0.83$	1.04	$84.03\pm 0.12$	1.54
SCAFFOLD	$71.93\pm 1.14$	85.26	$48.53\pm 0.34$	18.40	$87.36\pm 0.37$	2.08	$84.13\pm 0.20$	3.08

Table 1: Accuracy and communication cost comparison across datasets.

Figure 3: t-SNE visualization of per-client class centroids based on penultimate features from 10 clients on CIFAR-10. Colors denote classes and marker shapes indicate clients.

Panel of Table 2 — Table 2: Per-client round wall-clock time for local computation, normalized to FedAvg on each dataset. Bars share a $0$ – $4\times$ scale, and vertical ticks mark the FedAvg baseline ( $1\times$ ).

Noise type	Noise level	CIFAR-10	CIFAR-100	SVHN	FEMNIST
Laplace	$\epsilon=2.0$	$74.69\pm 0.97$	$49.09\pm 0.59$	$88.49\pm 0.34$	$85.84\pm 0.19$
	$\epsilon=1.0$	$74.42\pm 0.97$	$47.85\pm 0.58$	$88.88\pm 0.40$	$85.76\pm 0.19$
	$\epsilon=0.5$	$74.14\pm 0.47$	$47.30\pm 0.82$	$88.80\pm 0.38$	$85.29\pm 0.19$
	$\epsilon=0.2$	$69.83\pm 0.97$	$45.11\pm 1.02$	$87.07\pm 0.41$	$84.77\pm 0.31$
Gaussian	$\delta=0.05$	$73.02\pm 0.84$	$46.18\pm 0.39$	$87.88\pm 0.59$	$85.27\pm 0.08$
	$\delta=0.01$	$71.07\pm 0.64$	$45.56\pm 1.02$	$86.67\pm 0.80$	$84.88\pm 0.30$

Table 3: Performance of ASFL with Pixel DP in accuracy (%). The radius of threat model,

R=0.1

Figure 4: Ablation study on selection of activation sharing layers. The left and right y-axes represent the communication cost and the validation accuracy, respectively. The x-axis denotes the layer index.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	CIFAR-10	CIFAR-100	SVHN	FEMNIST
	(ResNet18)	(MobileNetV2)	(ResNet20)	(CNN)
Learning rate	0.1	0.05	0.1	0.01
Batch size	64	64	32	20
Weight decay	1e-4	5e-4	1e-4	1e-4
LR decay rounds	100, 150	100, 150	100, 150	–
Communication rounds	200	200	200	200

Appendix

Table 4: Training settings used for the main comparison in Table 1 . Each column corresponds to the dataset–backbone pair used in the main experiments.

Method	Hyperparameter	CIFAR-10	CIFAR-100	SVHN	FEMNIST
		(ResNet18)	(MobileNetV2)	(ResNet20)	(CNN)
FedProx	$\mu$ (proximal coefficient)	0.001	0.001	0.001	0.001
FedProto	$\lambda$ (prototype regularization coefficient)	1.0	1.0	1.0	1.0
MOON	$\mu$ (contrastive coefficient), $T$ (temperature)	1.0, 0.5	1.0, 0.5	0.1, 0.5	1.0, 0.5

Appendix

Table 5: Method-specific hyperparameters used in the main comparison.

Dataset	Backbone	No. of candidate positions	Selected layer block
CIFAR-10	ResNet18	9	4
CIFAR-100	MobileNetV2	9	2
SVHN	ResNet20	10	5
FEMNIST	CNN	4	2

Appendix

Table 6: Candidate activation-sharing positions for each dataset–backbone pair. The number of candidate positions excludes the input boundary but includes the final classifier position.

	$l=0$	$l=1$	$l=2$	$l=3$	$l=4$	$l=5$	$l=6$	$l=7$	$l=8$	$l=9$
ASFL	$79.6\pm\text{0.71}$	$76.74\pm\text{0.47}$	$76.87\pm\text{0.38}$	$77.38\pm\text{0.26}$	$75.24\pm\text{0.03}$	$73.85\pm\text{0.31}$	$70.14\pm\text{0.52}$	$69.38\pm\text{0.44}$	$65.68\pm\text{0.21}$	$66.00\pm\text{0.58}$
ASFL + Laplace	$67.59\pm\text{0.95}$	$69.77\pm\text{1.32}$	$67.05\pm\text{0.87}$	$70.53\pm\text{0.95}$	$69.83\pm\text{0.97}$	$70.86\pm\text{0.40}$	$67.93\pm\text{0.38}$	$66.81\pm\text{0.95}$	$65.09\pm\text{0.37}$	$65.34\pm\text{0.95}$
ASFL + Gaussian	$66.59\pm\text{1.20}$	$68.54\pm\text{0.18}$	$68.91\pm\text{1.31}$	$69.63\pm\text{0.74}$	$68.55\pm\text{0.70}$	$70.16\pm\text{0.79}$	$68.10\pm\text{0.46}$	$67.60\pm\text{0.48}$	$65.48\pm\text{0.29}$	$65.27\pm\text{0.94}$
Laplace Drop	$11.47\downarrow$	$6.97\downarrow$	$9.82\downarrow$	$6.85\downarrow$	$5.41\downarrow$	$2.99\downarrow$	$2.21\downarrow$	$2.57\downarrow$	$0.59\downarrow$	$0.66\downarrow$
Gaussian Drop	$12.47\downarrow$	$8.20\downarrow$	$7.96\downarrow$	$7.75\downarrow$	$6.69\downarrow$	$3.69\downarrow$	$2.04\downarrow$	$1.78\downarrow$	$0.20\downarrow$	$0.73\downarrow$

Appendix

Table 7: Impact of activation sharing layer on performance of CIFAR-10 with Laplace (

\epsilon=0.2

) and Gaussian (

\delta=0.001,\epsilon=0.5

). Laplace Drop and Gaussian Drop denote the accuracy reduction (%) of ASFL + Laplace and ASFL + Gaussian relative to the ASFL baseline, respectively.

	$l=0$	$l=1$	$l=2$	$l=3$	$l=4$	$l=5$	$l=6$	$l=7$	$l=8$	$l=9$
ASFL	$50.88\pm\text{0.78}$	$49.91\pm\text{0.40}$	$49.12\pm\text{0.73}$	$47.81\pm\text{0.24}$	$40.95\pm\text{0.95}$	$31.82\pm\text{1.13}$	$32.44\pm\text{2.18}$	$39.21\pm\text{0.85}$	$44.08\pm\text{1.22}$	$47.23\pm\text{0.58}$
ASFL + Laplace	$47.16\pm\text{0.59}$	$46.20\pm\text{0.54}$	$45.11\pm\text{1.02}$	$43.91\pm\text{0.92}$	$37.48\pm\text{0.91}$	$29.92\pm\text{0.57}$	$28.60\pm\text{1.45}$	$38.06\pm\text{0.83}$	$41.50\pm\text{0.80}$	$44.80\pm\text{0.48}$
ASFL + Gaussian	$45.84\pm\text{1.00}$	$46.40\pm\text{0.24}$	$45.19\pm\text{0.05}$	$43.48\pm\text{0.94}$	$39.61\pm\text{2.41}$	$30.90\pm\text{0.57}$	$29.75\pm\text{0.75}$	$38.03\pm\text{0.05}$	$41.17\pm\text{0.48}$	$46.34\pm\text{0.17}$
Laplace Drop	$3.72\downarrow$	$3.71\downarrow$	$4.01\downarrow$	$3.9\downarrow$	$3.47\downarrow$	$1.90\downarrow$	$3.84\downarrow$	$1.15\downarrow$	$2.58\downarrow$	$2.43\downarrow$
Gaussian Drop	$5.04\downarrow$	$3.71\downarrow$	$3.93\downarrow$	$4.33\downarrow$	$1.34\downarrow$	$0.92\downarrow$	$2.69\downarrow$	$1.18\downarrow$	$2.91\downarrow$	$0.89\downarrow$

Appendix

Table 8: Impact of activation sharing layer on performance of CIFAR-100 with Laplace (

\epsilon=0.2

) and Gaussian (

\delta=0.001,\epsilon=0.5

). Laplace Drop and Gaussian Drop denote the accuracy reduction (%) of ASFL + Laplace and ASFL + Gaussian relative to the ASFL baseline, respectively.

	$l=0$	$l=1$	$l=2$	$l=3$	$l=4$	$l=5$	$l=6$	$l=7$	$l=8$	$l=9$	$l=10$
ASFL	$89.20\pm\text{0.16}$	$89.37\pm\text{0.10}$	$89.11\pm\text{0.03}$	$89.40\pm\text{0.18}$	$89.34\pm\text{0.11}$	$89.12\pm\text{0.10}$	$88.62\pm\text{0.28}$	$89.43\pm\text{0.55}$	$88.39\pm\text{0.15}$	$86.61\pm\text{0.41}$	$86.02\pm\text{0.59}$
ASFL + Laplace	$19.66\pm\text{0.09}$	$51.61\pm\text{24.99}$	$19.61\pm\text{0.02}$	$41.58\pm\text{31.07}$	$64.19\pm\text{31.53}$	$87.07\pm\text{0.41}$	$87.56\pm\text{0.40}$	$87.71\pm\text{0.88}$	$86.76\pm\text{0.65}$	$85.97\pm\text{0.30}$	$85.31\pm\text{0.90}$
ASFL + Gaussian	$22.73\pm\text{4.44}$	$32.99\pm\text{18.77}$	$19.60\pm\text{0.01}$	$54.13\pm\text{25.37}$	$64.76\pm\text{31.94}$	$85.95\pm\text{0.81}$	$85.82\pm\text{0.92}$	$86.67\pm\text{0.90}$	$88.00\pm\text{0.25}$	$85.73\pm\text{0.73}$	$85.36\pm\text{1.23}$
Laplace Drop	$69.54\downarrow$	$37.76\downarrow$	$69.50\downarrow$	$47.82\downarrow$	$25.15\downarrow$	$2.05\downarrow$	$1.06\downarrow$	$1.72\downarrow$	$1.63\downarrow$	$0.64\downarrow$	$0.71\downarrow$
Gaussian Drop	$66.47\downarrow$	$56.38\downarrow$	$69.51\downarrow$	$35.27\downarrow$	$25.48\downarrow$	$3.17\downarrow$	$2.80\downarrow$	$2.76\downarrow$	$0.39\downarrow$	$0.88\downarrow$	$0.66\downarrow$

Appendix

Table 9: Impact of activation sharing layer on performance of SVHN with Laplace (

\epsilon=0.2

) and Gaussian (

\delta=0.001,\epsilon=0.5

). Laplace Drop and Gaussian Drop denote the accuracy reduction (%) of ASFL + Laplace and ASFL + Gaussian relative to the ASFL baseline, respectively.

	$l=0$	$l=1$	$l=2$	$l=3$	$l=4$
ASFL	$86.38\pm\text{0.17}$	$86.27\pm\text{0.57}$	$86.01\pm\text{0.20}$	$85.56\pm\text{0.35}$	$85.76\pm\text{0.24}$
ASFL + Laplace	$83.66\pm\text{0.14}$	$83.25\pm\text{0.51}$	$84.77\pm\text{0.31}$	$85.03\pm\text{0.08}$	$85.18\pm\text{0.25}$
ASFL + Gaussian	$83.66\pm\text{0.19}$	$83.55\pm\text{0.40}$	$84.96\pm\text{0.08}$	$85.05\pm\text{0.24}$	$85.18\pm\text{0.16}$
Laplace Drop	$2.72\downarrow$	$3.02\downarrow$	$1.24\downarrow$	$0.53\downarrow$	$0.58\downarrow$
Gaussian Drop	$2.72\downarrow$	$2.72\downarrow$	$1.05\downarrow$	$0.51\downarrow$	$0.58\downarrow$

Appendix

Table 10: Impact of activation sharing layer on performance of FEMNIST with Laplace (

\epsilon=0.2

) and Gaussian (

\delta=0.001,\epsilon=0.5

). Laplace Drop and Gaussian Drop denote the accuracy reduction (%) of ASFL + Laplace and ASFL + Gaussian relative to the ASFL baseline, respectively.

Figure 5: Optimization-based feature inversion attack on shared activations under Gaussian perturbation. The top row shows the original private input images for reference, the middle row visualizes the corresponding perturbed shared activations used for inversion, and the bottom row shows the reconstructed images.

Strategy	CIFAR-10 (ResNet18)	FEMNIST (CNN)
Random	$\mathbf{75.25}\pm 0.32$	$\mathbf{84.49}\pm 0.59$
Replication	$74.34\pm 0.76$	$84.11\pm 0.38$
Over-sampling	$71.95\pm 0.83$	$84.39\pm 0.30$

Appendix

Table 11: Ablation study on activation pool construction. We report validation accuracy (%) for three strategies for handling received remote activations: Random, Replication, and Over-sampling.

Shared	CIFAR-10		CIFAR-100		SVHN		FEMNIST
Batches	Acc (%) $\uparrow$	Comm (MB) $\downarrow$	Acc (%) $\uparrow$	Comm (MB) $\downarrow$	Acc (%) $\uparrow$	Comm (MB) $\downarrow$	Acc (%) $\uparrow$	Comm (MB) $\downarrow$
$8\times$	74.15 $\pm$ 1.17	64	53.27 $\pm$ 0.86	32	89.06 $\pm$ 0.27	8	84.82 $\pm$ 0.28	1.92
$4\times$	76.01 $\pm$ 0.50	32	52.34 $\pm$ 1.12	16	88.99 $\pm$ 0.49	4	84.59 $\pm$ 0.21	0.96
$2\times$	75.81 $\pm$ 0.70	16	51.18 $\pm$ 0.69	8	89.26 $\pm$ 0.65	2	84.61 $\pm$ 0.07	0.48
$1\times$	75.80 $\pm$ 0.55	8	50.24 $\pm$ 0.54	4	89.29 $\pm$ 0.31	1	84.62 $\pm$ 0.28	0.24

Appendix

Table 12: Ablation study on activation sharing quantity (the number of shared activation batches).

Method	$\alpha=0.05$	$\alpha=0.5$
FedAvg	$58.93\pm 0.3$	$85.77\pm 0.3$
FedProx	$56.16\pm 0.1$	$86.20\pm 0.1$
FedProto	$59.44\pm 1.0$	$86.86\pm 0.2$
MOON	$58.69\pm 1.1$	$86.58\pm 0.1$
SCAFFOLD	$55.37\pm 1.8$	$87.59\pm 0.1$
FedPVR	$58.45\pm 0.9$	$86.27\pm 0.1$

Appendix

Table 13: Validation accuracy (%) on CIFAR-10 with different Dirichlet concentration parameters. Smaller

\alpha

corresponds to stronger data heterogeneity.

Method	Validation accuracy (%)
FedAvg	$67.70\pm 0.33$
ASFL (without remote activations)	$66.23\pm 0.41$
ASFL (with remote activations)	$75.27\pm 0.88$

Appendix

Table 14: Ablation study on activation sharing on CIFAR-10. Both ASFL variants use the same partial training procedure with 16 tail update steps per tail round.

Explore similar work

FedPLT: Scalable, Resource-Efficient, and Heterogeneity-Aware Federated Learning via Partial Layer Training
May 4, 2026Ahmad Dabaja, Rachid El-AzouziHeterogeneous Federated Learning Federated Learning
Laboratoire Informatique d’Avignon (LIA), Avignon University, France
FedXDS: Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning
Jun 30, 2026Maximilian Andreas Hoefler, Karsten Mueller, Wojciech SamekFederated Learning Heterogeneity
Fraunhofer Heinrich Hertz Institute, Berlin, Germany · Technical University of Berlin, Berlin, Germany · Berlin Institute for the Foundations of Learning and Data (BIFOLD)
Subspace Optimization for Efficient Federated Learning under Heterogeneous Data
Apr 28, 2026Shuchen Zhu, Zhengyang Huang, Yuqi Xu +1Heterogeneous Federated Learning Federated Learning
Peking University · Beihang University

May 4, 2026, cs.DCJ/K move · Enter open · S save

FedPLT: Scalable, Resource-Efficient, and Heterogeneity-Aware Federated Learning via Partial Layer Training

Ahmad Dabaja, Rachid El-Azouzi

Laboratoire Informatique d’Avignon (LIA), Avignon University, France

Heterogeneous Federated Learning Federated Learning Heterogeneity Parameter-Efficient Fine-Tuning Methods Resource-Efficient

Open paper AskPDF arXiv

Federated Learning (FL) has gained significant attention in distributed machine learning by enabling collaborative model training across decentralized system while preserving data privacy. Although extensive research has addressed statistical data heterogeneity, FL still faces several challenges, including high communication and computation overheads and severe device heterogeneity, which require further investigation. Prior work has addressed these issues through sub-model training and partial parameter training. However, such methods often suffer from inconsistent parameter distributions across clients, inaccurate global loss estimation, and increased bias and variance. Guided by our empirical analysis, we propose FedPLT (Federated Learning with Partial Layer Training), an innovative and structured partial parameter training approach that exhibits training behavior similar to full model training while assigning client-specific portions of the model according to their communication and computational capabilities. In addition, we evaluate the performance of FedPLT when combined with optimal client sampling under communication constraints. We show that this integration improves FL performance by reducing sampling variance under the same communication budget. Through extensive experiments, we demonstrate that FedPLT achieves performance comparable to, or even surpassing, that of full-model training (i.e., FedAvg), while requiring significantly fewer trainable parameters per client. Moreover, FedPLT outperforms existing methods in highly heterogeneous environments, effectively adapts to client resource constraints, and reduces the number of straggling clients. In particular, FedPLT reduces the number of trainable parameters by 71%-82% while achieving performance on par with full-model training.

Abstract

Explore similar work

FedPLT: Scalable, Resource-Efficient, and Heterogeneity-Aware Federated Learning via Partial Layer Training

FedXDS: Leveraging Model Attribution Methods to counteract Data Heterogeneity in Federated Learning

Subspace Optimization for Efficient Federated Learning under Heterogeneous Data