Split Federated Unlearning

Jia Chen; Yecheng Lv; Xinrong Hu; Mingdi Huang; Jun Bai

cs.LGOct 27, 2026

Split Federated Unlearning

Authors: Jia Chen, Yecheng Lv, Xinrong Hu, Mingdi Huang, Jun Bai

Organizations: School of Computer Science and Artificial Intelligence, Wuhan Textile University, Wuhan, China · Engineering Research Center of Hubei Province for Clothing Information, Wuhan Textile University, Wuhan, China · Australian Institute for Machine Learning (AIML), Adelaide University, Adelaide, Australia

Abstract

Split federated learning (SFL) combines partitioned model execution with federated aggregation to support collaborative learning across distributed Web and edge environments. However, removing learned contributions in SFL is challenging due to cross-partition coupling and historical influence propagation through dual-side aggregation. To the best of our knowledge, we provide the first systematic formulation of Split Federated Unlearning (SFU), encompassing client-wise and class-wise removal, counterfactual retraining, and partition-specific update constraints. We theoretically characterize historical influence propagation and mode-dependent unlearning feasibility, revealing structural limitations of restricted model updates. As an initial solution, we propose SURE, integrating dual-importance estimation, selective parameter perturbation, and importance-guided recovery. Extensive experiments with existing unlearning methods adapted to SFL reveal substantial forgetting-utility trade-offs across model architectures, data distributions, and update modes. SURE achieves competitive agreement with mode-matched retraining at modest additional cost, while further analyses highlight the impact of split position and update constraints on unlearning behavior. These findings establish SFU as a distinct research problem and provide a foundation for unlearning in partitioned federated systems.

Figures & tables

Figure 1. Overview of Split Federated Unlearning (SFU). (a) SFL jointly trains client-side models and server-side replicas, with independent aggregation of both partitions. (b) SFU considers client-wise and class-wise removal, using full retained-data retraining as the counterfactual reference under different model update modes. (c) SFU involves cross-partition coupling, historical influence propagation, and mode-dependent unlearning feasibility.

Figure 2. Overview of SURE. (a) Dual-importance estimation identifies parameters relevant to forgotten and retained data across both model partitions. (b) Mode-aware perturbation targets one layer within the permitted partitions. (c) Retained-data recovery employs importance-guided gradient modulation and dual-side aggregation while preserving frozen partitions.

Client-only

Server-only

Joint

Method

Ret.

\uparrow

For.

Test

\uparrow

MIA

\downarrow

Ret.

\uparrow

For.

Test

\uparrow

MIA

\downarrow

Ret.

\uparrow

For.

Test

\uparrow

MIA

\downarrow

MNIST / LeNet

Retrain

88.24_{\scriptscriptstyle\pm 1.48}

96.97_{\scriptscriptstyle\pm 0.35}

89.36_{\scriptscriptstyle\pm 0.37}

44.26_{\scriptscriptstyle\pm 0.68}

0.00

88.35_{\scriptscriptstyle\pm 2.24}

0.00_{\scriptscriptstyle\pm 0.00}

80.17_{\scriptscriptstyle\pm 0.47}

53.75_{\scriptscriptstyle\pm 0.92}

0.00

89.71_{\scriptscriptstyle\pm 2.41}

0.00_{\scriptscriptstyle\pm 0.00}

81.07_{\scriptscriptstyle\pm 0.74}

52.33_{\scriptscriptstyle\pm 0.98}

0.00

Fine-tune

88.03_{\scriptscriptstyle\pm 1.26}

96.97_{\scriptscriptstyle\pm 4.96}

\mathbf{89.62}_{\scriptscriptstyle\pm 2.18}

46.70_{\scriptscriptstyle\pm 2.98}

0.73

92.04_{\scriptscriptstyle\pm 1.63}

4.74_{\scriptscriptstyle\pm 2.32}

83.42_{\scriptscriptstyle\pm 1.25}

52.93_{\scriptscriptstyle\pm 3.92}

3.13

\mathbf{93.87}_{\scriptscriptstyle\pm 1.52}

17.03_{\scriptscriptstyle\pm 14.96}

86.17_{\scriptscriptstyle\pm 1.62}

53.40_{\scriptscriptstyle\pm 1.98}

6.84

FedEraser

86.86_{\scriptscriptstyle\pm 2.58}

96.54_{\scriptscriptstyle\pm 3.65}

87.86_{\scriptscriptstyle\pm 2.75}

44.07_{\scriptscriptstyle\pm 3.27}

0.88

81.95_{\scriptscriptstyle\pm 2.72}

0.00_{\scriptscriptstyle\pm 0.00}

73.72_{\scriptscriptstyle\pm 1.35}

49.74_{\scriptscriptstyle\pm 3.62}

4.22

81.12_{\scriptscriptstyle\pm 1.21}

0.00_{\scriptscriptstyle\pm 0.00}

73.02_{\scriptscriptstyle\pm 1.65}

47.83_{\scriptscriptstyle\pm 2.25}

5.29

FedRecovery

86.00_{\scriptscriptstyle\pm 0.86}

96.54_{\scriptscriptstyle\pm 0.43}

87.14_{\scriptscriptstyle\pm 0.94}

50.59_{\scriptscriptstyle\pm 5.77}

2.81

86.06_{\scriptscriptstyle\pm 1.23}

95.67_{\scriptscriptstyle\pm 0.21}

\mathbf{87.14}_{\scriptscriptstyle\pm 1.62}

46.70_{\scriptscriptstyle\pm 5.26}

28.00

86.29_{\scriptscriptstyle\pm 1.23}

95.24_{\scriptscriptstyle\pm 0.43}

\mathbf{86.60}_{\scriptscriptstyle\pm 0.34}

48.18_{\scriptscriptstyle\pm 4.62}

27.09

Table 1. Client-wise unlearning under extreme non-IID partitions (mean

\pm

SD, %). AG (Avg. Gap)

\downarrow

is the mean absolute distance of Ret., For., Test, and MIA from mode-matched Retrain (percentage points), computed from reported means. Ret./Test

\uparrow

: higher is better. Bold marks the best non-reference Ret./Test or AG within each dataset–model and update mode.

Client-only

Server-only

Joint

Method

Ret.

\uparrow

For.

Test

\uparrow

MIA

\downarrow

Ret.

\uparrow

For.

Test

\uparrow

MIA

\downarrow

Ret.

\uparrow

For.

Test

\uparrow

MIA

\downarrow

MNIST / LeNet

Retrain

99.23_{\scriptscriptstyle\pm 0.02}

99.43_{\scriptscriptstyle\pm 0.25}

98.74_{\scriptscriptstyle\pm 0.21}

49.46_{\scriptscriptstyle\pm 4.13}

0.00

99.11_{\scriptscriptstyle\pm 0.10}

0.00_{\scriptscriptstyle\pm 0.00}

88.96_{\scriptscriptstyle\pm 0.35}

50.33_{\scriptscriptstyle\pm 1.44}

0.00

99.12_{\scriptscriptstyle\pm 0.14}

0.00_{\scriptscriptstyle\pm 0.00}

89.01_{\scriptscriptstyle\pm 0.31}

49.34_{\scriptscriptstyle\pm 3.14}

0.00

Fine-tune

99.19_{\scriptscriptstyle\pm 0.06}

99.43_{\scriptscriptstyle\pm 0.25}

97.81_{\scriptscriptstyle\pm 0.23}

48.99_{\scriptscriptstyle\pm 0.77}

0.36

\mathbf{99.52}_{\scriptscriptstyle\pm 0.10}

92.67_{\scriptscriptstyle\pm 1.14}

98.34_{\scriptscriptstyle\pm 0.23}

46.79_{\scriptscriptstyle\pm 3.92}

26.50

\mathbf{99.55}_{\scriptscriptstyle\pm 0.06}

93.10_{\scriptscriptstyle\pm 1.55}

98.37_{\scriptscriptstyle\pm 0.22}

48.57_{\scriptscriptstyle\pm 2.26}

25.92

SalUn

97.92_{\scriptscriptstyle\pm 0.30}

86.93_{\scriptscriptstyle\pm 2.49}

96.32_{\scriptscriptstyle\pm 0.33}

47.09_{\scriptscriptstyle\pm 5.48}

4.65

98.37_{\scriptscriptstyle\pm 0.94}

7.47_{\scriptscriptstyle\pm 2.17}

88.91_{\scriptscriptstyle\pm 0.47}

50.47_{\scriptscriptstyle\pm 0.21}

2.10

99.38_{\scriptscriptstyle\pm 0.26}

7.04_{\scriptscriptstyle\pm 2.17}

89.61_{\scriptscriptstyle\pm 0.54}

50.24_{\scriptscriptstyle\pm 0.21}

2.20

MoDe

99.18_{\scriptscriptstyle\pm 0.10}

99.57_{\scriptscriptstyle\pm 0.43}

98.71_{\scriptscriptstyle\pm 0.27}

47.75_{\scriptscriptstyle\pm 1.84}

0.48

96.46_{\scriptscriptstyle\pm 0.44}

0.00_{\scriptscriptstyle\pm 0.00}

86.93_{\scriptscriptstyle\pm 0.88}

48.75_{\scriptscriptstyle\pm 2.36}

1.57

96.70_{\scriptscriptstyle\pm 0.46}

0.00_{\scriptscriptstyle\pm 0.00}

87.01_{\scriptscriptstyle\pm 0.97}

47.68_{\scriptscriptstyle\pm 1.66}

1.52

Table 2. Class-wise unlearning under Dirichlet non-IID partitions (

\alpha=0.1

). Ret., For., Test, and MIA are mean

\pm

SD (%). AG (Avg. Gap)

\downarrow

is the mean absolute difference from mode-matched Retrain (pp), computed from reported metric means. Ret./Test

\uparrow

: higher is better; bold highlights the best non-reference Ret./Test means and lowest AG.

Figure 3. Impact of split position on client-wise unlearning with SURE on CIFAR-10/ResNet-18. Panels show (a) Test, (b) Forget, and (c) Retain accuracy across nine split positions. Colored curves represent Client-only, Server-only, and Joint updates, while dashed lines indicate mode-matched Retrain references.

Figure 4. Layer-wise CKA across nine split positions for client-wise unlearning on CIFAR-10/ResNet-18. The three panels correspond to Client-only, Server-only, and Joint updates. Colored curves represent different split positions, while the red dashed curve denotes Retrain. All CKA values are computed relative to Original.

Method	Ret. $\uparrow$	For.	Test $\uparrow$	MIA	AG $\downarrow$	Ret. $\uparrow$	For.	Test $\uparrow$	MIA	AG $\downarrow$	Ret. $\uparrow$	For.	Test $\uparrow$	MIA	AG $\downarrow$
	Client-only					Server-only					Joint
Retrain	$80.67_{\scriptscriptstyle\pm 1.96}$	$68.14_{\scriptscriptstyle\pm 1.08}$	$76.38_{\scriptscriptstyle\pm 1.67}$	$50.60_{\scriptscriptstyle\pm 2.01}$	0.00	$79.67_{\scriptscriptstyle\pm 1.48}$	$65.37_{\scriptscriptstyle\pm 3.33}$	$75.11_{\scriptscriptstyle\pm 0.82}$	$49.67_{\scriptscriptstyle\pm 1.24}$	0.00	$79.46_{\scriptscriptstyle\pm 2.47}$	$64.13_{\scriptscriptstyle\pm 2.32}$	$75.77_{\scriptscriptstyle\pm 1.86}$	$51.12_{\scriptscriptstyle\pm 0.23}$	0.00
A1	$80.99_{\scriptscriptstyle\pm 2.46}$	$66.00_{\scriptscriptstyle\pm 4.95}$	$76.53_{\scriptscriptstyle\pm 1.84}$	$51.13_{\scriptscriptstyle\pm 1.35}$	0.79	$81.38_{\scriptscriptstyle\pm 1.20}$	$71.80_{\scriptscriptstyle\pm 4.75}$	$77.00_{\scriptscriptstyle\pm 1.34}$	$50.51_{\scriptscriptstyle\pm 4.29}$	2.72	$80.06_{\scriptscriptstyle\pm 5.67}$	$69.74_{\scriptscriptstyle\pm 1.42}$	$75.86_{\scriptscriptstyle\pm 3.97}$	$49.22_{\scriptscriptstyle\pm 3.29}$	2.05
A2	$\mathbf{81.23}_{\scriptscriptstyle\pm 2.50}$	$67.83_{\scriptscriptstyle\pm 3.74}$	$\mathbf{76.89}_{\scriptscriptstyle\pm 2.09}$	$50.87_{\scriptscriptstyle\pm 2.15}$	0.41	$81.03_{\scriptscriptstyle\pm 2.31}$	$70.93_{\scriptscriptstyle\pm 1.17}$	$76.12_{\scriptscriptstyle\pm 1.59}$	$49.54_{\scriptscriptstyle\pm 1.05}$	2.02	$\mathbf{82.20}_{\scriptscriptstyle\pm 2.46}$	$70.59_{\scriptscriptstyle\pm 2.44}$	$77.35_{\scriptscriptstyle\pm 1.36}$	$52.12_{\scriptscriptstyle\pm 1.86}$	2.95
A3	$79.40_{\scriptscriptstyle\pm 1.48}$	$65.24_{\scriptscriptstyle\pm 3.92}$	$75.77_{\scriptscriptstyle\pm 1.72}$	$50.24_{\scriptscriptstyle\pm 2.92}$	1.29	$79.62_{\scriptscriptstyle\pm 5.68}$	$70.61_{\scriptscriptstyle\pm 3.16}$	$74.70_{\scriptscriptstyle\pm 4.84}$	$51.28_{\scriptscriptstyle\pm 2.75}$	1.83	$79.81_{\scriptscriptstyle\pm 3.19}$	$70.60_{\scriptscriptstyle\pm 2.71}$	$75.87_{\scriptscriptstyle\pm 2.44}$	$49.71_{\scriptscriptstyle\pm 2.95}$	2.08
A4	$78.75_{\scriptscriptstyle\pm 1.55}$	$64.91_{\scriptscriptstyle\pm 4.49}$	$75.20_{\scriptscriptstyle\pm 2.00}$	$51.05_{\scriptscriptstyle\pm 2.10}$	1.70	$81.68_{\scriptscriptstyle\pm 2.62}$	$73.00_{\scriptscriptstyle\pm 3.31}$	$76.77_{\scriptscriptstyle\pm 1.49}$	$50.47_{\scriptscriptstyle\pm 1.20}$	3.03	$82.13_{\scriptscriptstyle\pm 1.54}$	$71.21_{\scriptscriptstyle\pm 3.82}$	$77.37_{\scriptscriptstyle\pm 1.18}$	$49.74_{\scriptscriptstyle\pm 3.21}$	3.18

Table 3. SURE ablation on CIFAR-10/ResNet-18 for client-wise unlearning (Dirichlet

\alpha=0.1

, split 2). A0 denotes full SURE; A1–A7 are described in Section 5.4 . Results are mean

\pm

SD (%). Ret. and Test (

\uparrow

) denote predictive utility, while For. and MIA are interpreted relative to mode-matched Retrain. AG (Avg. Gap) (

\downarrow

, pp) is the mean absolute difference from Retrain across the four metrics. Bold marks the highest non-Retrain utility or lowest gap.

	Client-only		Server-only		Joint
Method	Comm. $\downarrow$	FLOPs $\downarrow$	Comm. $\downarrow$	FLOPs $\downarrow$	Comm. $\downarrow$	FLOPs $\downarrow$
CIFAR-10 / CNN
Fine-tune	53.3	333	26.5	218	53.3	379
FedEraser	53.5	333	26.5	218	53.5	379
FedRecovery	0.189	0.00707	0.161	0.00707	0.189	0.00707
SalUn	59.1	371	29.6	244	59.1	422

Table 4. Communication and computation costs for client-wise unlearning under extreme non-IID partitions. Communication is in GB (

10^{9}

bytes) and computation in TFLOPs (

10^{12}

operations); lower is better.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Client-only

Server-only

Joint

Method

Ret.

\uparrow

For.

Test

\uparrow

MIA

\downarrow

Ret.

\uparrow

For.

Test

\uparrow

MIA

\downarrow

Ret.

\uparrow

For.

Test

\uparrow

MIA

\downarrow

CIFAR-10 / CNN

Retrain

84.62_{\scriptscriptstyle\pm 0.18}

81.72_{\scriptscriptstyle\pm 0.80}

81.09_{\scriptscriptstyle\pm 0.23}

49.53_{\scriptscriptstyle\pm 1.50}

0.00

85.32_{\scriptscriptstyle\pm 0.38}

81.50_{\scriptscriptstyle\pm 0.82}

81.74_{\scriptscriptstyle\pm 0.24}

50.42_{\scriptscriptstyle\pm 0.52}

0.00

84.88_{\scriptscriptstyle\pm 0.72}

81.28_{\scriptscriptstyle\pm 0.82}

81.54_{\scriptscriptstyle\pm 0.63}

49.80_{\scriptscriptstyle\pm 1.24}

0.00

Fine-tune

85.23_{\scriptscriptstyle\pm 1.10}

84.14_{\scriptscriptstyle\pm 1.56}

82.01_{\scriptscriptstyle\pm 0.60}

50.79_{\scriptscriptstyle\pm 1.00}

1.30

86.06_{\scriptscriptstyle\pm 1.32}

83.48_{\scriptscriptstyle\pm 1.24}

81.85_{\scriptscriptstyle\pm 0.92}

51.73_{\scriptscriptstyle\pm 1.16}

1.04

87.23_{\scriptscriptstyle\pm 0.24}

85.46_{\scriptscriptstyle\pm 1.36}

81.98_{\scriptscriptstyle\pm 0.37}

48.82_{\scriptscriptstyle\pm 1.52}

1.99

FedEraser

82.00_{\scriptscriptstyle\pm 2.00}

82.38_{\scriptscriptstyle\pm 0.43}

79.87_{\scriptscriptstyle\pm 1.47}

50.48_{\scriptscriptstyle\pm 1.52}

1.36

82.76_{\scriptscriptstyle\pm 1.30}

80.40_{\scriptscriptstyle\pm 0.48}

79.91_{\scriptscriptstyle\pm 1.42}

50.42_{\scriptscriptstyle\pm 1.32}

1.37

75.12_{\scriptscriptstyle\pm 2.42}

71.59_{\scriptscriptstyle\pm 0.33}

72.65_{\scriptscriptstyle\pm 2.35}

51.37_{\scriptscriptstyle\pm 1.36}

7.48

FedRecovery

84.84_{\scriptscriptstyle\pm 0.36}

85.02_{\scriptscriptstyle\pm 0.72}

81.55_{\scriptscriptstyle\pm 0.27}

50.49_{\scriptscriptstyle\pm 0.84}

1.24

84.84_{\scriptscriptstyle\pm 1.25}

85.02_{\scriptscriptstyle\pm 0.42}

81.56_{\scriptscriptstyle\pm 1.12}

50.43_{\scriptscriptstyle\pm 0.94}

1.05

84.84_{\scriptscriptstyle\pm 0.36}

85.02_{\scriptscriptstyle\pm 0.53}

81.56_{\scriptscriptstyle\pm 0.73}

50.49_{\scriptscriptstyle\pm 0.35}

1.12

Appendix

Table 5. Client-wise unlearning under IID partitions (mean

\pm

SD, %). AG (Avg. Gap)

\downarrow

is the mean absolute distance of Ret., For., Test, and MIA from mode-matched Retrain (percentage points), computed from reported means. Ret./Test

\uparrow

: higher is better. Bold marks the best non-reference Ret./Test or AG within each dataset–model and update mode.

Figure 5. t-SNE visualization of client-wise unlearning on CIFAR-10/ResNet-18 at split 1 (left) and split 8 (right). Each row compares Client-only, Server-only, and Joint updates. Blue and green denote retained-training and non-member samples, respectively, while pink and yellow indicate forgotten samples before and after SURE.

Figure 6. t-SNE visualization of class-wise unlearning on CIFAR-10/ResNet-18 under Joint updates. Panels (a) and (b) show client cut-layer and server pre-classifier representations, respectively, each comparing Original, Retrain, and SURE. Colors denote classes, and black triangles mark forgotten class-0 samples.

Mode	Metric	10–1	10–2	10–5	20–2	50–5
Client-only	Retain (%)	81.23 $\pm$ 2.50	81.78 $\pm$ 0.51	72.37 $\pm$ 0.83	72.42 $\pm$ 0.83	61.19 $\pm$ 0.76
	Forget (%)	67.83 $\pm$ 3.74	69.05 $\pm$ 1.22	70.98 $\pm$ 1.28	74.67 $\pm$ 0.66	51.69 $\pm$ 0.36
	Test (%)	76.89 $\pm$ 2.09	76.87 $\pm$ 0.77	72.21 $\pm$ 0.44	71.41 $\pm$ 0.71	58.33 $\pm$ 0.59
	MIA (%)	50.87 $\pm$ 2.15	51.62 $\pm$ 0.55	51.80 $\pm$ 0.47	51.19 $\pm$ 0.79	50.25 $\pm$ 0.94
	Avg. Gap $\downarrow$	0.41	0.91	1.96	0.82	0.81
Server-only	Retain (%)	81.16 $\pm$ 1.99	82.36 $\pm$ 0.52	75.18 $\pm$ 0.37	75.22 $\pm$ 0.71	57.63 $\pm$ 0.35

Appendix

Table 6. SURE with multiple forgetting clients on CIFAR-10/ResNet-18 under Dirichlet

\alpha=0.1

. A configuration

T

–

F

denotes

T

total clients and

F

forgetting clients. Avg. Gap (pp) is computed against the corresponding Retrain reference.

Figure 7. Sensitivity of SURE to perturbation protection and recovery coefficients on CIFAR-10/ResNet-18 under Dirichlet $\alpha=0.1$ . (a) Avg. Gap as the perturbation protection coefficient $\gamma$ varies under Client-only, Server-only, and Joint updates. (b) Avg. Gap over recovery coefficients $(\beta_{R},\beta_{F})$ for the three update modes. Lower is better. Circles and red boxes mark the best setting within each mode.

Federated Unlearning

Federated unlearning removes a client’s or record’s influence from a collaboratively trained model while preserving useful knowledge.

See all 29 papers

pFedUL: Layer-Aware Federated Unlearning for Personalized Federated Learning
Jun 15, 2026Zhuodong Liu, Xiangyu Li, Zhihao ZhangFederated Unlearning Federated Learning
1Saunders College of Business, Rochester Institute of Technology, Rochester, NY 14623, USA · School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA 30332, USA
Rethinking Federated Unlearning via the Lens of Memorization
May 23, 2026Jiaheng Wei, Yanjun Zhang, He Zhang +5Federated Unlearning Neural Network Memorization
Royal Melbourne Institute of Technology · Griffith University · Swinburne University of Technology
Not All Retained Clients Are Equal: Client Selection for Federated Unlearning
Oct 22, 2026Thanh Linh Nguyen, Quoc-Viet PhamFederated Unlearning Client Selection in Federated Learning
School of Computer Science and Statistics, Trinity College Dublin, Ireland

Jun 15, 2026J/K move · Enter open · S save

pFedUL: Layer-Aware Federated Unlearning for Personalized Federated Learning

Zhuodong Liu, Xiangyu Li, Zhihao Zhang

1Saunders College of Business, Rochester Institute of Technology, Rochester, NY 14623, USA · School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA 30332, USA

Federated Unlearning Federated Learning

Open paper AskPDF arXiv

Federated unlearning (FU) enables the removal of specific data contributions from federated learning (FL) models to comply with regulations such as the General Data Protection Regulation (GDPR). However, most existing FU methods are designed for the FedAvg paradigm, where all clients share a single global model. In practice, personalized federated learning (pFL) methods such as FedPer, FedRep, Ditto, and FedBN have become widely adopted due to their superior handling of non-IID data. These methods decompose the model into shared global layers and client-specific personalized layers, fundamentally altering the semantics of unlearning, yet this setting has received little attention. We formalize FU under the pFL paradigm, identifying a tension between unlearning completeness on shared layers and personalization preservation for remaining clients. We then propose pFedUL, a layer-aware selective unlearning framework comprising three components: (1) gradient-based layer-wise contribution attribution that separately quantifies the target client's influence on shared and personalized parameters, (2) adaptive selective unlearning that applies differentiated forgetting strategies across layer types, and (3) a lightweight recalibration protocol enabling remaining clients to restore personalization with minimal overhead. We further introduce two new metrics, Personalization Preservation Score (PPS) and Cross-client Fairness Index (CFI), to evaluate pFL-specific unlearning quality. Experiments on CIFAR-10, CIFAR-100, and FEMNIST under varying non-IID settings indicate that pFedUL achieves unlearning effectiveness comparable to full retraining while maintaining an average of 97.3% personalized accuracy for remaining clients. Compared with six state-of-the-art FU methods adapted to the pFL setting, pFedUL consistently achieves superior personalization preservation.

Abstract

Federated Unlearning

Explore similar work

pFedUL: Layer-Aware Federated Unlearning for Personalized Federated Learning

Rethinking Federated Unlearning via the Lens of Memorization

Not All Retained Clients Are Equal: Client Selection for Federated Unlearning