Computation and Communication Efficient Federated Unlearning via On-server Gradient Conflict Mitigation and Expression

Minh-Duong Nguyen; Senura Hansaja; Le-Tuan Nguyen; Quoc-Viet Pham; Ken-Tye Yong; Nguyen H. Tran; Dung D. Le

cs.LGMar 14, 2026

Computation and Communication Efficient Federated Unlearning via On-server Gradient Conflict Mitigation and Expression

Authors: Minh-Duong Nguyen, Senura Hansaja, Le-Tuan Nguyen, Quoc-Viet Pham, Ken-Tye Yong, Nguyen H. Tran, Dung D. Le

Organizations: VinUniversity · University of Sydney · Trinity College Dublin

Abstract

Federated Unlearning (FUL) aims to remove specific participants' data contributions from a trained Federated Learning model, thereby ensuring data privacy and compliance with regulatory requirements. Despite its potential, progress in FUL has been limited due to several challenges, including the cross-client knowledge inaccessibility and high computational and communication costs. To overcome these challenges, we propose Federated On-server Unlearning (FOUL), a novel framework that comprises two key stages. The learning-to-unlearn stage serves as a preparatory learning phase, during which the model identifies and encodes the key features associated with the forget clients. This stage is communication-efficient and establishes the basis for the subsequent unlearning process. Subsequently, on-server knowledge aggregation phase aims to perform the unlearning process at the server without requiring access to client data, thereby preserving both efficiency and privacy. We introduce a new data setting for FUL, which enables a more transparent and rigorous evaluation of unlearning. To highlight the effectiveness of our approach, we propose a novel evaluation metric termed time-to-forget, which measures how quickly the model achieves optimal unlearning performance. Extensive experiments conducted on three datasets under various unlearning scenarios demonstrate that FOUL outperforms the Retraining in FUL. Moreover, FOUL achieves competitive or superior results with significantly reduced time-to-forget, while maintaining low communication and computation costs.

Figures & tables

Figure 1: Extension of the causal structure model to the multi-client setting. Following [ 28 ] , the causal factors

{\mathcal{Z}}_{K}

represent the domain-invariant characteristics. Consequently, when unlearning a client, removing

{\mathcal{Z}}_{K}

may reduce the amount of useful information necessary for predicting

{\mathcal{Y}}

. In contrast, unlearning the non-causal factors

{\mathcal{Z}}_{V,2}

primarily removes domain-specific information from client 2, while having a smaller impact on the information relevant to the classification of

{\mathcal{Y}}

. Dashed lines denote edges that may vary across clients or be absent in certain scenarios, solid lines indicate edges that remain invariant across all clients.

Figure 2: The learning on a single local client in the L2U stage. Featurizers are trained to disentangle knowledge into invariant and variant representations.

Figure 3: Unlearning stage with data-free on-server unlearning. Only the non-causal featurizer participates in this phase. The gradients of the non-causal featurizer from both retain and forget users are sent to the server for on-server gradient matching.

Group	Methods	PACS				TerraIncognita
Group	Methods	FA $(\downarrow)$	RA $(\uparrow)$	TA $(\uparrow)$	MIA $(\downarrow)$	FA $(\downarrow)$	RA $(\uparrow)$	TA $(\uparrow)$	MIA $(\downarrow)$
Retrain	Retrain	$70.51$	$82.84$	$77.45$	$50.02$	$30.64$	$42.41$	$38.94$	$50.71$
	FATS	$74.45$ ( $+3.94$ )	$80.91$ ( $-1.93$ )	$75.98$ ( $-1.47$ )	$55.72$ ( $+5.70$ )	$33.07$ ( $+2.43$ )	$40.96$ ( $-1.45$ )	$37.34$ ( $-1.60$ )	$60.81$ ( $+10.10$ )
	NoT	$73.24$ ( $+2.73$ )	$79.25$ ( $-3.59$ )	$75.28$ ( $-2.17$ )	$59.13$ ( $+9.11$ )	$32.26$ ( $+1.62$ )	$38.32$ ( $-4.09$ )	$36.08$ ( $-2.86$ )	$62.90$ ( $+12.19$ )
Appro. Unlearn.	FedCDP	$77.37$ ( $+6.86$ )	$78.13$ ( $-4.71$ )	$76.41$ ( $-1.04$ )	$72.26$ ( $+22.24$ )	$34.56$ ( $+3.92$ )	$38.79$ ( $-3.62$ )	$36.75$ ( $-2.19$ )	$81.91$ ( $+31.20$ )
	FedRecovery	$76.48$ ( $+5.97$ )	$76.97$ ( $-5.87$ )	$74.81$ ( $-2.64$ )	$75.24$ ( $+25.22$ )	$32.11$ ( $+1.47$ )	$37.49$ ( $-4.92$ )	$35.59$ ( $-3.35$ )	$84.03$ ( $+33.32$ )

Table 1: Comparison of methods on ResNet-18 with PACS and TerraIncognita datasets. The table reports Forget Accuracy (FA), Remaining Accuracy (RA), Testing Accuracy (TA), and Membership Inference Attack (MIA), with values in parentheses showing the difference from the Retrain baseline. For the FOUL method, (1) corresponds to the L2U phase, and (2) refers to the on-server gradient matching unlearning phase. Note that we bold results achieving the closest TA accuracy to Retrain.

Figure 4: Retraining from scratch with FedAvg without resetting model parameters on PACS dataset.

Figure 5: Retraining from scratch with FedAvg with resetting model parameters on PACS dataset.

Figure 6: Evaluations of FOUL on PACS dataset.

Figure 7: Comparison of communication cost in trade-off with Forget Accuracy (FA) and Remain Accuracy (RA) across different unlearning methods on PACS and TerraIncognita datasets.

Method	Params (M)	Comm. (MB)	Comp. (FLOPs)
	(M)	(MB)	(FLOPs)
Retrain	11.3	42.73	$5.81e^{16}$
FATS	11.3	42.73	$5.81e^{16}$
NoT	11.3	34.72	$3.38e^{16}$
FedCDP	11.3	42.73	$5.82e^{16}$
FedRecovery	11.3	42.73	$5.96e^{16}$

Table 2: Comparison of computational and communication characteristics of each method. Params denotes trainable parameters, Comm. refers to per-round communication overhead, and Comp. indicates local computation complexity.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

\displaystyle\mathcal{L}_{\textrm{L2U}}=\alpha_{\textrm{rec}}\mathcal{L}_{\textrm{rec}}+\alpha_{\textrm{iv}}\mathcal{L}_{\textrm{iv}}+\alpha_{\textrm{v}}\mathcal{L}_{\textrm{v}}.

Appendix

Algorithm 1 Pseudo algorithm for L2U stage of FOUL.

\Gamma_{\mathcal{R}}^{*},\Gamma_{\mathcal{F}}^{*}=\arg\min_{\Gamma_{\mathcal{R}},\Gamma_{\mathcal{F}}}(g^{(r)}_{\Gamma_{\mathcal{R}}}-g^{(r)}_{\Gamma_{\mathcal{F}}})\cdot g^{(r)}_{\textrm{FL}}+\sqrt{\kappa}\|g^{(r)}_{\textrm{FL}}\|\|g^{(r)}_{\Gamma_{\mathcal{R}}}-g^{(r)}_{\Gamma_{\mathcal{F}}}\|.

Appendix

Algorithm 2 Unlearning Stage via On-server Matching Gradient

Group	Methods	VLCS				OfficeHome
Group	Methods	FA $(\downarrow)$	RA $(\uparrow)$	TA $(\uparrow)$	MIA $(\downarrow)$	FA $(\downarrow)$	RA $(\uparrow)$	TA $(\uparrow)$	MIA $(\downarrow)$
Retrain	Retrain	$63.14$	$73.88$	$70.16$	$52.96$	$45.90$	$66.32$	$59.08$	$47.64$
	FATS	$66.51$ $(+3.37)$	$72.09$ $(-1.79)$	$68.62$ $(-1.54)$	$58.68$ $(+5.72)$	$48.73$ $(+2.83)$	$64.05$ $(-2.27)$	$\mathbf{58.13}$ $(-0.95)$	$57.46$ $(+9.82)$
	NoT	$65.52$ $(+2.38)$	$71.04$ $(-2.84)$	$67.83$ $(-2.33)$	$61.93$ $(+8.97)$	$47.57$ $(+1.67)$	$61.51$ $(-4.81)$	$56.15$ $(-2.93)$	$59.79$ $(+12.15)$
Appro. Unlearn.	FedCDP	$69.78$ $(+6.64)$	$69.95$ $(-3.93)$	$68.63$ $(-1.53)$	$74.56$ $(+21.60)$	$50.07$ $(+4.17)$	$61.83$ $(-4.49)$	$57.73$ $(-1.35)$	$78.06$ $(+30.42)$
	FedRecovery	$69.33$ $(+6.19)$	$68.79$ $(-5.09)$	$67.61$ $(-2.55)$	$77.34$ $(+24.38)$	$47.65$ $(+1.75)$	$60.81$ $(-5.51)$	$56.35$ $(-2.73)$	$80.41$ $(+32.77)$

Appendix

Table 3: Comparison of methods on ResNet-18 with VLCS and OfficeHome datasets. The table reports Forget Accuracy (FA), Remaining Accuracy (RA), Testing Accuracy (TA), and Membership Inference Attack (MIA), with values in parentheses showing the difference from the Retrain baseline. For the FOUL method, (1) corresponds to the L2U phase, and (2) refers to the on-server gradient matching unlearning phase. Note that we bold results achieving the closest TA accuracy to Retrain.

Panel of Figure 8 — Figure 8: Learning phase of FOUL under different learning coefficient settings.

Domain 1 (+90% )
Algorithm	+90%	+80%	-90%
Accuracy	89.5 $\pm$ 0.1	82.3 $\pm$ 0.5	73.2 $\pm$ 0.3
Invariant	0.0003 $\pm$ 10e-5	0.0002 $\pm$ 10e-5	0.0002 $\pm$ 10e-5
Domain 2 (+80% )
Accuracy	73.9 $\pm$ 0.1	82.3 $\pm$ 0.5	73.2 $\pm$ 0.3
Invariant	0.0003 $\pm$ 10e-5	0.0002 $\pm$ 10e-5	0.0002 $\pm$ 10e-5

Appendix

Table 4: Assessment of FOUL across different domains. The evaluation is conducted on the Colored-MNIST dataset, where each domain corresponds to images with varying degrees of correlation between color and label. Two aspects are assessed: (1) Accuracy: evaluating the model’s ability to predict labels in each domain using only the invariant representations; and (2) Invariance: measuring the disparity between the invariant representations

z^{i}_{K}

generated from a target domain

i

and those

z^{j}_{K}

obtained from the training domain

j

, to quantify the consistency of invariant features across domains.

Figure 9: Illustrations of the efficacy of invariant knowledge featurizer.

Panel of Figure 10 — Figure 10: Effects of different global learning rates and hypersphere radii on FOUL performance. For both PACS and Terra Incognita, the model achieves the best remaining accuracy (RA) and forgetting accuracy (FA) when $\eta=0.2$ and $\kappa=0.5$ .

Figure 11: Ablation test on different causal dimensionality of the model disentanglement.

Federated Unlearning

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

See all 27 papers

Jun 23, 2026

FedUP: One-Shot Federated Unlearning via Centroid-Guided Plug-in Filters

Federated unlearning (FU) is critical for complying with legal mandates like the right to be forgotten in decentralized systems, yet current methods face a persistent dilemma between non-target knowledge loss and high request latency. To resolve these issues, we propose FedUP, a one-shot federated unlearning framework utilizing lightweight pluggable filters that act as a "knowledge funnel" to screen out target data while preserving original model performance. By freezing original model parameters and training filters at the server side using differentially private (DP)-protected class centroid samples, FedUP bypasses the need for multi-round client-server communication and complex retraining, reducing unlearning latency from minutes to mere seconds. Additionally, the framework's pluggable architecture ensures inherent reversibility, enabling the seamless restoration of forgotten knowledge by simply removing the filters. Extensive experiments on diverse image and text tasks demonstrate that FedUP effectively reduces non-target knowledge loss and achieves superior unlearning precision and efficiency across various scenarios. Code is available at: https://github.com/suows/FedUP-code.

Feihong Nan, Zhengyi Zhong, Pan Wang +4

1National Key Laboratory of Big Data and Decision, National University of Defense Technology, China.

Federated Unlearning Machine Unlearning Federated Learning

Ask PDF

Jun 15, 2026

pFedUL: Layer-Aware Federated Unlearning for Personalized Federated Learning

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.

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

Ask PDF

Apr 6, 2026

Forgetting to Witness: Efficient Federated Unlearning and Its Visible Evaluation

With the increasing importance of data privacy and security, federated unlearning has emerged as a novel research field dedicated to ensuring that federated learning models no longer retain or leak relevant information once specific data has been deleted. In this paper, to the best of our knowledge, we propose the first complete pipeline for federated unlearning, which includes a federated unlearning approach and an evaluation framework. Our proposed federated unlearning approach ensures high efficiency and model accuracy without the need to store historical data.It effectively leverages the knowledge distillation model alongside various optimization mechanisms. Moreover, we propose a framework named Skyeye to visualize the forgetting capacity of federated unlearning models. It utilizes the federated unlearning model as the classifier integrated into a Generative Adversarial Network (GAN). Afterward, both the classifier and discriminator guide the generator in generating samples. Throughout this process, the generator learns from the classifier's knowledge. The generator then visualizes this knowledge through sample generation. Finally, the model's forgetting capability is evaluated based on the relevance between the deleted data and the generated samples. Comprehensive experiments are conducted to illustrate the effectiveness of the proposed federated unlearning approach and the corresponding evaluation framework.

Houzhe Wang, Xiaojie Zhu, Chi Chen

School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China · Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China · King Abdullah University of Science and Technology, Thuwal, Kingdom of Saudi Arabia

Federated Unlearning Machine Unlearning Evaluation

Ask PDF

Abstract

Federated Unlearning

Explore similar work

FedUP: One-Shot Federated Unlearning via Centroid-Guided Plug-in Filters

pFedUL: Layer-Aware Federated Unlearning for Personalized Federated Learning

Forgetting to Witness: Efficient Federated Unlearning and Its Visible Evaluation