One-Shot Federated Model Editing for Device Dynamics at the Edge

Youming Tao; Weijun Yin; Huanyi Xie; Yanqing Yao; Dongxiao Yu; Xiuzhen Cheng; Di Wang; Falko Dressler

cs.LGOct 26, 2026

One-Shot Federated Model Editing for Device Dynamics at the Edge

Authors: Youming Tao, Weijun Yin, Huanyi Xie, Yanqing Yao, Dongxiao Yu, Xiuzhen Cheng, Di Wang, Falko Dressler

Organizations: School of Electrical Engineering and Computer Science, TU Berlin, 10623 Berlin, Germany · School of Cyber Science and Technology, Beihang University, Beijing 100191, China · Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing, Beihang University, Beijing 100083, China · Division of Computer, Electrical and Mathematical Sciences and Engineering, King Abdullah University of Science and Technology, Saudi Arabia ·

Abstract

Standard federated learning (FL) requires persistent device connectivity across multiple communication rounds, making it unsuitable for dynamic environments such as vehicular networks, where devices frequently join and leave. One-shot FL offers a promising alternative by completing training in a single communication round. However, one-shot FL suffers from model solidification and cannot readily adapt to subsequent client dynamics, including new client contributions and the withdrawal of previous contributions. We therefore consider exact model editing, which removes withdrawn contributions through unlearning or incorporates newly contributed ones through incremental learning without retraining from scratch. We identify that efficient editing fundamentally hinges on learning stability. To this end, we propose F-SOSA, the first stable one-shot FL algorithm that enables rapid and exact model editing. F-SOSA leverages sub-sampling of local models with carefully tuned rate for stable aggregation and performs curvature-aware model fusion to preserve utility. We prove that F-SOSA achieves vanishing training error and reasonably small generalization error bounds. Building on this stability, we further develop two single-round editing algorithms: F-SOSA-U for unlearning and F-SOSA-I for incremental learning. They ensure that the edited model, along with all intermediate algorithmic states, is statistically indistinguishable from retraining while incurring only an expected $O(ρ)$ fraction of its cost, where $ρ$ quantifies stability. Experiments demonstrate that our approach matches retraining performance while reducing cumulative time and communication overhead by at least 36% and 22%, respectively, under realistic urban mobility simulations.

Figures & tables

Structure	Storage / Communication	Curvature Structure
Full	$O(d_{\ell}^{2})$	Full
Diagonal	$O(d_{\ell})$	Diagonal only
K-FAC	$O(i_{\ell}^{2}+o_{\ell}^{2})$	Kronecker-factorized

TABLE I: Structural complexity of representative per-client Fisher representations for a layer with

d_{\ell}=o_{\ell}i_{\ell}

parameters.

\boldsymbol{\bar{W}}_{\mathcal{P}}\leftarrow\boldsymbol{\bar{W}}_{\mathcal{P}}+\tfrac{1}{P}\big(\boldsymbol{\tilde{W}}_{i^{+}}-\boldsymbol{\tilde{W}}_{i^{-}}\big),

Algorithm 2 F-SOSA-U

\boldsymbol{\bar{W}}_{\mathcal{P}}\leftarrow\boldsymbol{\bar{W}}_{\mathcal{P}}+\frac{1}{P}\big(\boldsymbol{\tilde{W}}_{i^{+}}-\boldsymbol{\tilde{W}}_{i^{-}}\big),

Algorithm 3 F-SOSA-I

	$\displaystyle\bar{\boldsymbol{W}}^{\prime}$	$\displaystyle\leftarrow\bar{\boldsymbol{W}}_{\mathcal{P}}+\frac{1}{P}\left(\sum_{i\in S_{+}}\widetilde{\boldsymbol{W}}_{i}-\sum_{i\in S_{-}}\widetilde{\boldsymbol{W}}_{i}\right),$
	$\displaystyle\boldsymbol{F}^{\prime}$	$\displaystyle\leftarrow\boldsymbol{F}_{\mathcal{P}}+\sum_{i\in S_{+}}\boldsymbol{F}_{i}-\sum_{i\in S_{-}}\boldsymbol{F}_{i},$
	$\displaystyle\boldsymbol{b}^{\prime}$	$\displaystyle\leftarrow\boldsymbol{b}_{\mathcal{P}}+\sum_{i\in S_{+}}\boldsymbol{F}_{i}\widetilde{\boldsymbol{W}}_{i}-\sum_{i\in S_{-}}\boldsymbol{F}_{i}\widetilde{\boldsymbol{W}}_{i}.$

Algorithm 4 F-SOSA-Batch : Paired Joint Editing

Dataset	Metrics	$\alpha=0.1$				$\alpha=0.5$				$\alpha=0.8$
Dataset	Metrics	FedFisher	FedLPA	Dense	F-SOSA	FedFisher	FedLPA	Dense	F-SOSA	FedFisher	FedLPA	Dense	F-SOSA
MNIST	Accuracy (%)	9 0.35	87.90	85.92	91.08	9 5.98	94.21	96.06	95.43	9 6.08	92.82	96.93	95.80
	Time (s)	43.11	24.59	187.96	3 4.63	43.76	26.61	190.54	3 4.59	43.15	29.02	193.58	3 5.56
	Comm. (MB)	35.30	793.31	3 3.76	18.46	35.30	793.31	3 3.76	18.46	35.30	793.31	3 3.76	18.46
	Time Efficiency	2.10	3.57	0.52	2 .63	2.19	3.54	0.57	2 .76	2.23	3.20	0.58	2 .69
	Comm. Efficiency	2 .56	0.11	2.55	4.93	2.72	0.12	2 .85	5.17	2.72	0.12	2 .87	5.19

TABLE II: Performance comparison of F-SOSA and baselines across MNIST, SVHN, and CIFAR-10 under varying Dirichlet heterogeneity

\alpha

. Best results are in bold; second-best results are underlined.

Metrics	FedFisher	FedLPA	Dense	F-SOSA
Accuracy (%)	76.94	70.18	7 5.97	75.71
Time (s)	115.89	18.69	138.57	6 7.07
Comm. (MB)	554.12	1968.70	5 28.83	289.68
Time Efficiency	0.66	3.75	0.64	1 .13
Comm. Efficiency	0.139	0.036	0 .144	0.261

TABLE III: Performance comparison on FEMNIST under a real-world non-IID setting. Best results are in bold; second-best results are underlined.

Method

FEMNIST

CIFAR-10

Accuracy (%)

Precision (%)

AUC (%)

Accuracy (%)

Precision (%)

AUC (%)

FedErasesr

50.01_{\pm 0.02}

51.08_{\pm 3.23}

50.01_{\pm 0.02}

48.99_{\pm 0.78}

40.75_{\pm 7.83}

48.99_{\pm 0.78}

FedRR

50.45_{\pm 0.49}

50.37_{\pm 0.41}

50.45_{\pm 0.49}

45.17_{\pm 2.33}

42.29_{\pm 2.87}

45.17_{\pm 2.33}

FATS-CU

50.47_{\pm 0.25}

50.33_{\pm 0.18}

50.47_{\pm 0.25}

53.54_{\pm 0.88}

61.29_{\pm 2.99}

53.54_{\pm 0.88}

Retrain

50.45_{\pm 0.21}

50.31_{\pm 0.15}

50.45_{\pm 0.21}

51.64_{\pm 1.67}

52.04_{\pm 2.09}

51.64_{\pm 1.67}

F-SOSA-U

50.92_{\pm 0.33}

50.63_{\pm 0.23}

50.92_{\pm 0.33}

51.63_{\pm 1.80}

52.17_{\pm 2.72}

51.63_{\pm 1.80}

TABLE IV: Membership Inference Attack (MIA) results on unlearned global models from different methods on FEMNIST and CIFAR-10.

Fig. 1: Effect of the stability level

\rho

on F-SOSA’s accuracy and time/ communication overhead across datasets.

Panel of Figure 9 — Fig. 2: Effect of local epoch number $K$ on performance. Top-left: MNIST, top-right: SVHN, bottom-left: FEMNIST, bottom-right: CIFAR-10.

Panel of Figure 10 — Fig. 3: Performance of F-SOSA-I/U vs. baselines on CIFAR-10. Columns: IL (left), UL (right). Rows: test accuracy (top), time cost (middle), communication cost (bottom).

Panel of Figure 11 — Fig. 4: Effect of stability parameter $\rho$ on accuracy and efficiency. Top-left: MNIST, top-right: SVHN, bottom-left: FEMNIST, bottom-right: CIFAR-10.

Fig. 5: The Luxembourg map and the region considered.

Panel of Figure 13 — Fig. 6: Test accuracy (left) and cumulative cost (right) of F-SOSA-I/U vs. Retrain under simulator-generated model editing requests. Rows: MNIST, SVHN, FEMNIST, CIFAR-10 (top to bottom).

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Metrics	$\alpha=0.1$				$\alpha=0.5$				$\alpha=0.8$
Dataset	Metrics	FedFisher	FedLPA	Dense	F-SOSA	FedFisher	FedLPA	Dense	F-SOSA	FedFisher	FedLPA	Dense	F-SOSA
CIFAR10 (Resnet18)	Accuracy (%)	46.60	4 5.18	28.04	35.17	69.66	6 8.41	57.48	65.98	72.61	7 0.55	63.92	69.87
	Time (s)	3 61.02	2493.05	479.82	278.22	3 62.14	2355.83	474.14	210.34	3 65.50	1713.55	473.76	208.27
	Comm. (MB)	8 82.16	8066.21	895.13	462.44	8 82.16	8066.21	895.13	462.44	8 82.16	8066.21	895.13	462.44
	Time Efficiency	0.129	0.018	0.058	0 .126	0 .192	0.029	0.121	0.314	0 .199	0.041	0.135	0.335
	Comm. Efficiency	0 .053	0.006	0.031	0.076	0 .079	0.008	0.064	0.143	0 .082	0.009	0.071	0.151

Appendix

TABLE V: Performance comparison of F-SOSA and baselines across CIFAR-10 and BDD100K under varying Dirichlet heterogeneity

\alpha

. Best results are in bold; second-best results are underlined.

Clients	Method	MNIST			SVHN			FEMNIST			CIFAR10
Clients	Method	Acc. (%)	T-save (%)	Comm. ( $\times$ )	Acc. (%)	T-save (%)	Comm. ( $\times$ )	Acc. (%)	T-save (%)	Comm. ( $\times$ )	Acc. (%)	T-save (%)	Comm. ( $\times$ )
(a) Incremental learning
20	Retrain	96.33	0.00	1.00	7 3.07	0.00	1.00	27.06	0.00	1.00	6 8.51	0.00	1.00
	Fine-tuning	9 5.74	79.68	0 .78	74.67	81.48	0 .78	21.65	69.13	0 .78	67.86	78.89	0 .78
	F-SOSA-I	95.62	2 0.83	0.75	72.82	3 7.84	0.75	2 5.64	1 9.67	0.75	69.34	4 5.26	0.75
50	Retrain	9 4.11	0.00	1.00	6 6.61	0.00	1.00	30.22	0.00	1.00	5 7.11	0.00	1.00

Appendix

TABLE VI: Scalability over 15 incremental-learning or unlearning requests. We report mean accuracy (Acc.), mean time-cost saving (T-save), and mean communication cost normalized to Retrain (Comm.; Retrain

=1.00

). T-save is relative to Retrain; higher is better for Acc. and T-save, while lower is better for Comm. Best results are in bold; second-best results are underlined.

Fig. 7: mAP

{}_{50\text{-}95}

(left) and cumulative cost (right) of F-SOSA-I/U vs. Retrain under simulator-generated model editing requests on the real driving dataset BDD100K.

Dataset	Model	Broadcast	Selected Clients			Unselected Clients	Round Total
		Downlink	Model Up.	Diag Up.	K-FAC Up.	Uplink	Diag/K-FAC
MNIST	Two-Layer CNN	1.61	8.04	8.04	8.81	0	17.69/18.46
SVHN	Two-Layer CNN	3.58	17.92	17.92	18.81	0	39.42/40.31
CIFAR10	Two-Layer CNN	3.58	17.92	17.92	18.81	0	39.42/40.31
CIFAR10	ResNet18-Pretrained	42.63	213.13	213.13	206.68	0	468.89/462.44
FEMNIST	Two-Layer CNN	25.18	125.91	125.91	138.59	0	277.00/289.68

Appendix

TABLE VII: Detailed communication costs (MB) of F-SOSA for

M=20

and

\rho=0.5

, where

M

is the total number of clients and

\rho

is the client selection ratio. The downlink corresponds to one broadcast of the global model to all active clients. Model Up., Diag Up., and K-FAC Up. denote the aggregate uplink communication of the selected clients for local models, diagonal Fisher information, and Kronecker-Factored Approximate Curvature (K-FAC) Fisher information, respectively. Unselected clients incur no uplink communication. The final column reports the total communication costs of the Diag and K-FAC variants.

Request	Broadcast	Selected Clients		Unselected Clients	Round Total
	Downlink	Model Up.	Fisher Up.	Uplink
Initial	1.61/1.61	4.82/4.82	4.82/4.82	0.00/0.00	11.25/11.25
12+	0.00/1.61	0.00/4.82	0.00/4.82	0.00/0.00	0.00/11.25
13+	1.61/1.61	5.63/5.63	5.63/5.63	0.00/0.00	12.87/12.87
14+	0.00/1.61	0.00/5.63	0.00/5.63	0.00/0.00	0.00/12.87
15+	1.61/1.61	6.43/6.43	6.43/6.43	0.00/0.00	14.47/14.47

Appendix

TABLE VIII: Detailed communication costs (MB) of sequential editing on MNIST with

\rho=0.5

. Each communication entry is reported as F-SOSA-I/U versus Retrain, with the first value corresponding to F-SOSA-I/U and the second to Retrain. Model Up. and Fisher Up. denote the communication costs of uploading the model and the approximate Fisher information, respectively. In the Request column, the number identifies the affected client, while

+

and

-

denote client addition and client removal, respectively.

Request	Broadcast	Selected Clients		Unselected Clients	Round Total
	Downlink	Model Up.	Fisher Up.	Uplink
Initial	3.58/3.58	10.75/10.75	11.28/11.28	0.00/0.00	25.61/25.61
12+	3.58/3.58	10.75/10.75	11.28/11.28	0.00/0.00	25.61/25.61
13+	3.58/3.58	12.55/12.55	13.16/13.16	0.00/0.00	29.29/29.29
14+	3.58/3.58	12.55/12.55	13.16/13.16	0.00/0.00	29.29/29.29
15+	3.58/3.58	14.34/14.34	15.04/15.04	0.00/0.00	32.96/32.96

Appendix

TABLE IX: Detailed communication costs (MB) of sequential editing on SVHN with

\rho=0.5

+

and

-

denote client addition and client removal, respectively.

Request	Broadcast	Selected Clients		Unselected Clients	Round Total
	Downlink	Model Up.	Fisher Up.	Uplink
Initial	25.18/25.18	75.55/75.55	83.15/83.15	0.00/0.00	183.88/183.88
12+	0.00/25.18	0.00/75.55	0.00/83.15	0.00/0.00	0.00/183.88
13+	25.18/25.18	88.14/88.14	97.01/97.01	0.00/0.00	210.33/210.33
14+	0.00/25.18	0.00/88.14	0.00/97.01	0.00/0.00	0.00/210.33
15+	25.18/25.18	100.73/100.73	110.87/110.87	0.00/0.00	236.78/236.78

Appendix

TABLE X: Detailed communication costs (MB) of sequential editing on FEMNIST with

\rho=0.5

+

and

-

denote client addition and client removal, respectively.

Request	Broadcast	Selected Clients		Unselected Clients	Round Total
	Downlink	Model Up.	Fisher Up.	Uplink
Initial	3.58/3.58	10.75/10.75	11.28/11.28	0.00/0.00	25.61/25.61
12+	3.58/3.58	10.75/10.75	11.28/11.28	0.00/0.00	25.61/25.61
13+	3.58/3.58	12.55/12.55	13.16/13.16	0.00/0.00	29.29/29.29
14+	0.00/3.58	0.00/12.55	0.00/13.16	0.00/0.00	0.00/29.29
15+	3.58/3.58	14.34/14.34	15.04/15.04	0.00/0.00	32.96/32.96

Appendix

TABLE XI: Detailed communication costs (MB) of sequential editing on CIFAR10 with

\rho=0.5

+

and

-

denote client addition and client removal, respectively.

Request	Broadcast	Selected Clients		Unselected Clients	Round Total
	Downlink	Model Up.	Fisher Up.	Uplink
Initial	42.49/42.49	127.48/127.48	127.20/127.20	0.00/0.00	297.17/297.17
12+	0.00/42.49	0.00/127.48	0.00/127.20	0.00/0.00	0.00/297.17
13+	42.49/42.49	148.73/148.73	148.40/148.40	0.00/0.00	339.62/339.62
14+	0.00/42.49	0.00/148.73	0.00/148.40	0.00/0.00	0.00/339.62
15+	42.49/42.49	169.97/169.97	169.60/169.60	0.00/0.00	382.06/382.06

Appendix

TABLE XII: Detailed communication costs (MB) of sequential editing on BDD100K with

\rho=0.5

+

and

-

denote client addition and client removal, respectively.

Federated Learning

Federated learning trains shared models across data holders without requiring them to pool their raw data.

See all 951 papers

Jul 9, 2026

FedOPAL: One-Shot Federated Learning via Analytic Visual Prompt Tuning

With the widespread deployment of basic models in edge intelligence, communication bandwidth has become a core bottleneck restricting the scalability of federated learning. Although one-shot federated learning alleviates this problem by minimizing communication rounds, existing iterative fine-tuning or knowledge distillation methods still face challenges such as high server-side computational costs and hyperparameter sensitivity. Analytical federated learning achieves efficient gradientfree aggregation using least-squares closed-form solutions, but in environments with non-independent and identically distributed data, its static feature assumptions fail, leading to feature manifold misalignment and severely impairing model performance. To address this contradiction, this paper proposes the FedOPAL framework. This framework adapts the visual prompts as feature rectifiers, actively correcting the feature distribution of heterogeneous data to a linearly separable space by applying local proximal constraints, thereby satisfying the theoretical assumptions of analytical federated learning. Experimental results show that FedOPAL not only significantly outperforms the original analytical methods on several benchmarks, but also achieves accuracy comparable to state-of-the-art iterative methods while maintaining zero server-side training costs, providing a new engineering paradigm for efficient collaboration of large models on the edge.

Lingyu Qiu, Daniela Annunziata, Stefano Izzo +2

Department of Mathematics and Applications “R. Caccioppoli”, University of Naples Federico II, Italy

Federated Learning Aggregation Non-IID Federated Learning Federated Learning

Ask PDF

Jul 9, 2026

Collate: Collaborative Neural Network Learning for Latency-Critical Edge Systems

Federated Learning (FL) empowers multiple clients to collaboratively learn a model, enlarging the training data of each client for high accuracy while protecting data privacy. However, when deploying FL in real-time edge systems, the heterogeneity of devices among systems has a severe impact on the performance of the inferred model. Existing optimizations on FL focus on improving the training efficiency but fail to speed up inference, especially when there is a latency constraint. In this work, we propose Collate, a novel training framework that collaboratively learns heterogeneous models to meet the latency constraints of multiple edge systems simultaneously. We design a dynamic zeroizing-recovering method to adjust each local model architecture for high accuracy under its latency constraint. A proto-corrected federated aggregation scheme is also introduced to aggregate all heterogeneous local models, satisfying the latency constraint of different systems with only one training process and maintaining high accuracy. Extensive experiments indicate that, compared to state-of-the-art methods and under a latency constraint, our extended models can improve the accuracy by 1.96% on average, and our shrunk models can also obtain a 3.09% accuracy improvement on average, with almost no extra training overhead. The related codes and data will be available at https://github.com/ntuliuteam/Collate

Shuo Huai, Di Liu, Hao Kong +5

School of Computer Science and Engineering, Nanyang Technological University, Singapore · HP-NTU Digital Manufacturing Corporate Lab, Nanyang Technological University, Singapore · HP Inc., Palo Alto, California, USA

Efficient Neural Network Inference Federated Learning Aggregation Edge Inference Federated Learning Heterogeneous Federated Learning

Ask PDF

Oct 26, 2026

Efficient Exact Unlearning for Federated Zeroth-Order Optimization

Federated zeroth-order optimization (FedZOO) reduces the memory and communication costs of collaborative training for edge intelligence by avoiding backpropagation while communicating only random seeds and scalar coefficients. However, existing methods lack efficient mechanisms for exact unlearning to fully remove the influence of training data when users request deletion. A key insight is that total variation stability of the learning algorithm creates overlap between model distributions before and after data deletion, allowing exact unlearning to reuse computation through coupling. Based on this observation, we jointly design two stable zeroth-order learning algorithms and their corresponding exact-unlearning procedures: FedZOO-S obtains stability through client and mini-batch subsampling, enabling FedZOUL-S to perform efficient unlearning by preserving earlier updates and restarting from the first affected step. FedZOO-NS further combines subsampling with Gaussian noise to increase overlap between update distributions; FedZOUL-NS exploits this overlap through rejection--reflection coupling to reuse affected updates when the coupling is accepted. We prove exact unlearning for both algorithms: the model and saved state have the same joint distribution as retraining with the corresponding learner on retained data. We further establish stability and convergence guarantees for the learning algorithms and derive upper bounds on expected unlearning cost. Experiments demonstrate effective unlearning with competitive model utility and up to

56.1\%

less cumulative unlearning time than re-optimization. Compared with first-order baselines, our methods reduce communication overhead by up to approximately four orders of magnitude.

Weijun Yin, Yanqing Yao, Falko Dressler +2

School of Cyber Science and Technology, Beihang University, Beijing 100191, China · Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing, Beihang University, Beijing 100083, China · School of Electrical Engineering and Computer Science, TU Berlin, 10623 Berlin, Germany +1

Federated Unlearning Federated Learning

Ask PDF

Abstract

Federated Learning

Explore similar work

FedOPAL: One-Shot Federated Learning via Analytic Visual Prompt Tuning

Collate: Collaborative Neural Network Learning for Latency-Critical Edge Systems

Efficient Exact Unlearning for Federated Zeroth-Order Optimization