Pith. sign in

REVIEW 3 major objections 7 minor 44 references

Forgetting a client in personalized federated learning needs layer-aware treatment of shared versus private parameters, and pFedUL can match retraining-level forgetting while keeping about 97% of remaining clients' personalized accuracy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Layer-aware selective unlearning for personalized FL matches near-retrain forgetting while retaining ~97% personalized accuracy for remaining clients across four pFL architectures.

T0 review reviewed 2026-07-15 challenge →

load-bearing objection Solid empirical paper that correctly names a real pFL unlearning gap and ships usable metrics; the headline 6.3% PPS win is partly inflated by giving recalibration only to pFedUL. the 3 major comments →

arxiv 2606.16304 v1 pith:OXYWRURJ submitted 2026-06-15 cs.LG

pFedUL: Layer-Aware Federated Unlearning for Personalized Federated Learning

Zhuodong Liu , Xiangyu Li , Zhihao Zhang This is my paper
classification cs.LG
keywords federated learningfederated unlearningpersonalized federated learningmachine unlearninglayer-aware unlearningright to be forgottenFisher informationpersonalization preservation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Standard federated unlearning assumes one shared global model. Personalized federated learning breaks that assumption by splitting parameters into shared layers everyone updates and private layers each client keeps. Removing one client then creates a real trade-off: you must erase that client's influence from the shared layers without wrecking the personalization remaining clients already depend on. This paper formalizes that trade-off and offers pFedUL, a three-stage method that scores each shared layer by how much the departing client shaped it, aggressively corrects only the high-influence layers, discards the target's private parameters, and lets remaining clients briefly retune their private parts. On image benchmarks under non-IID splits it approaches full retraining on forgetting metrics while preserving roughly 97.3% of remaining clients' personalized accuracy, beating adapted prior unlearning methods on personalization and fairness and doing so with far less communication.

Core claim

Under personalized federated learning, unlearning is not uniform parameter scrubbing: completeness on shared layers and preservation of remaining clients' personalization pull in opposite directions. A layer-aware pipeline that attributes the target client's influence via diagonal Fisher scores, applies selective constrained ascent only to high-contribution shared layers, and finishes with 1–2 epochs of frozen-shared local recalibration can approximate the retrained shared model while largely restoring each remaining client's personalized performance.

What carries the argument

Layer-wise contribution score I^{c_t}_l from exponential-moving-average diagonal Fisher statistics: the fraction of total per-layer sensitivity attributable to the target client, used to gate adaptive selective unlearning intensity and leave low-score shared layers untouched.

Load-bearing premise

The method assumes that a running average of squared gradients stored during training is a faithful enough map of which shared layers encode the departing client's data, so correcting only those layers and briefly retuning private heads is enough without seeing the client's raw data again.

What would settle it

On the same CIFAR/FEMNIST non-IID setups, check whether unlearning only the high-Fisher layers still drives membership-inference accuracy near 0.5 while low-Fisher-only or random-layer unlearning leaves high membership-inference accuracy; if high-Fisher selection fails to separate forgetting from utility, the attribution premise fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Privacy-compliant removal of a client can be offered inside common personalized FL designs (FedPer, FedRep, Ditto, FedBN) without a full federation retrain.
  • Remaining clients need only a short private recalibration after a shared-layer edit, not multi-round re-aggregation.
  • PPS and CFI become usable evaluation dimensions alongside membership-inference accuracy when personalization matters.
  • Selective, contribution-weighted forgetting scales better as more clients request simultaneous or sequential removal than uniform gradient ascent.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same shared-versus-private split that makes personalization work is exactly what makes naive unlearning unfair across remaining clients; any future FU baseline that ignores that split will systematically understate personalization damage.
  • Storing per-client Fisher diagonals is lighter than full gradient histories but still leaks distributional cues; privacy-preserving storage of those scores is a natural next compliance requirement.
  • If the concentration of client influence into a few layers holds for transformers and LoRA-style adapters, the same selective principle could transfer to federated foundation-model unlearning without full parameter rewrites.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The paper formalifies federated unlearning (FU) under personalized FL (pFL) architectures that split parameters into shared layers θ_s and client-specific layers θ_p_i (FedPer, FedRep, Ditto, FedBN). It identifies a completeness–preservation trade-off and proposes pFedUL: (1) EMA diagonal-Fisher layer-wise contribution scores I^{c_t}_l (Eqs. 4–6), (2) adaptive selective constrained gradient ascent on high-contribution shared layers only (Eqs. 7–11), and (3) 1–2 epochs of frozen-shared local recalibration of remaining clients’ personalized parameters (Eqs. 12–13). Two pFL-specific metrics, PPS and CFI (Eqs. 14–15), are introduced. On CIFAR-10/100 and FEMNIST, pFedUL reports MIA within ~1–1.8 points of Retrain, average PPS ≈ 0.973, ~6.3% PPS gains over six adapted SOTA FU methods, and an 8.4× wall-clock speedup with zero global aggregation rounds (Tables 4–6, 8).

Significance. The gap between mainstream FU (FedAvg-centric) and widely deployed pFL is real and under-addressed; formalizing the shared/personalized tension and evaluating four pFL backbones is a useful contribution. PPS and CFI are sensible, non-tautological complements to MIA/remaining accuracy. Strengths include multi-dataset, multi-architecture tables with mean±std over five seeds; an ablation (Fig. 2); a controlled attribution-validity experiment (Table 7) showing high-I layers drive forgetting; sensitivity and multi-client/challenging-scenario analyses (Fig. 3, Table 9); and an explicit limitations discussion of Fisher storage privacy. If the comparative claims hold under matched post-processing, the work would be a solid empirical advance for regulatory-compliant pFL.

major comments (3)
  1. §4.1 and Table 3 state that all adapted baselines (including ZeroFU/Mimir/FUSED) receive no post-unlearning recalibration, while pFedUL always applies Stage 3 (E_r=2, frozen θ̂_s). Fig. 2 shows that removing recalibration alone drops PPS from 0.975 to 0.948—about half the reported gap to ZeroFU-adapted (0.912 on CIFAR-10, Table 6). The abstract’s and §4.4’s claim of a ~6.3% PPS advantage over the best of six SOTA methods is therefore not cleanly attributable to Fisher-based selective correction. A matched-recalibration control (give every baseline the same E_r=2 head/BN/proximal fine-tune, and/or report pFedUL w/o Stage 3 against baselines w/o Stage 3 as the primary comparison) is load-bearing for the central comparative claim.
  2. The unlearning direction ĝ^{c_t}_l = F̂_{t,l} ⊙ (θ^s_l − θ^s_{l,init}) (Eq. 10) is a heuristic combining Fisher sensitivity with displacement from initialization. Remark 2 motivates Fisher as importance, but does not justify the (θ − θ_init) factor or show that the resulting ascent approximates the leave-one-client-out optimum in Eq. (3). Table 7 validates layer selection, not the update direction. Either a short leave-one-out / influence-function sanity check on a small model, or an ablation replacing (θ − θ_init) with stored target gradients / random signs under the same clip, is needed to support that Stage 2 (not only Stage 3) drives near-Retrain MIA.
  3. §3.3–3.4 and Algorithm 1 require per-client, per-layer EMA Fisher vectors stored for the full training trajectory. §5.2 acknowledges privacy risk but provides no quantification of storage (bytes vs. FedEraser checkpoints) or of how DP noise on F̂ (mitigation a) affects I^{c_t}_l ranking and downstream MIA/PPS. Given that the method’s privacy story is a primary motivation (GDPR), a small storage table and one DP-noised Fisher run on CIFAR-10/FedRep would make the practicality claim load-bearing rather than aspirational.
minor comments (7)
  1. Abstract and §1 claim “gradient-based” attribution in one place and “Fisher information-based” in others; align terminology with Eqs. (4)–(6).
  2. DOI line “http://doi.org/10.32604/cmc.2026.000000” and “Day Month Year” placeholders should be cleaned for the camera-ready version.
  3. Table 1 rightmost column is useful but dense; a binary “pFL-aware?” column would improve scannability.
  4. Eq. (15) CFI = 1 − std(preservation ratios) can go negative if dispersion > 1; state the practical range or clip, and note sensitivity for low-A^{pre}_i clients already flagged in Remark 3.
  5. Fig. 2 is described but numerical values for all four variants are only partially in text; a small companion table would help reproducibility.
  6. §4.1 sets η_u = 0.1 as “10× training LR” citing [31]; a one-sentence note that this is a design choice, not a theorem of [31], would avoid overclaim.
  7. References [17] and [35] share overlapping venue/year metadata with nearby entries; double-check bibliographic uniqueness (e.g., NoT vs. FUSED page ranges).

Circularity Check

0 steps flagged

No significant circularity: empirical systems paper whose claims rest on experimental measurements, not on results forced by definition or self-citation.

full rationale

pFedUL is a designed three-stage procedure (Fisher EMA attribution Eqs. 4–6, selective constrained ascent Eqs. 7–11, frozen-shared recalibration Eqs. 12–13) whose performance numbers (MIA near 0.5, average PPS 0.973, 6.3% PPS lift, 8.4× speedup) are obtained by running the algorithm on CIFAR-10/100 and FEMNIST and comparing against Retrain and six adapted baselines. PPS and CFI are ordinary ratio statistics of measured pre/post accuracies; they equal 1 only for the explicit Retrain upper bound by definition of the metric, not by construction of the unlearning map. Fisher scores are accumulated during ordinary training and used solely for layer selection/intensity; they are never fitted to the reported PPS or MIA targets. No uniqueness theorem, ansatz, or load-bearing self-citation reduces any central claim to its inputs. Experimental design choices (asymmetric recalibration of baselines) affect fairness of comparison but do not create definitional circularity in the derivation chain. The paper is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 3 invented entities

The central empirical claim rests on standard FL/pFL modeling assumptions, the established use of diagonal Fisher as parameter importance (EWC/influence-function literature), and a handful of hand-chosen unlearning hyper-parameters whose values are justified by sensitivity plots but remain free. No new physical entities are postulated; the invented objects are the algorithmic pipeline and two evaluation scores.

free parameters (6)
  • threshold sensitivity α = 0.5
    Controls which layers are high-contribution (τ = μ_I + α σ_I); default 0.5 chosen via trade-off plot in §4.7; directly affects which parameters are modified and therefore PPS/MIA.
  • unlearning step size η_u = 0.1
    Scales the constrained gradient-ascent update on high-contribution layers; set to 0.1 (10× training LR) by common unlearning practice, not derived.
  • clipping coefficient γ = 0.01
    Bounds per-layer parameter change to ~1% of layer norm; default 0.01 chosen to avoid catastrophic degradation.
  • Fisher EMA momentum β = 0.99
    Controls accumulation of squared gradients into ̂F; default 0.99 following Adam-style practice.
  • recalibration epochs E_r = 2
    Number of local epochs remaining clients run with frozen shared parameters; default 2 from diminishing-returns observation (PPS 0.948→0.975→0.977).
  • Ditto proximal λ = 0.1
    Regularization strength for Ditto personalization; fixed at 0.1 in experiments.
axioms (5)
  • domain assumption pFL models decompose into shared parameters θ^s aggregated across clients and client-specific personalized parameters θ^p_i (FedPer/FedRep heads, Ditto full local models, FedBN stats).
    Stated in §3.1 and Table 2; foundational to the problem formulation and to treating target θ^p_ct as simply deletable.
  • domain assumption Diagonal Fisher information (expected squared gradients) quantifies relative parameter importance of a client's data for each layer, justifying the contribution score I^{c_t}_l.
    Invoked in Remark 2 citing EWC and influence functions; the selective-unlearning strategy stands or falls on this proxy being sufficiently accurate.
  • domain assumption Historical per-client per-layer squared-gradient EMA can be stored on the server during training without violating FL privacy constraints enough to invalidate the method.
    Required by Stage 1 (Eq. 5); privacy caveats are discussed in §5.2 but the method assumes the storage is acceptable.
  • domain assumption Membership-inference attack accuracy near 0.5 plus remaining-client accuracy near retrain is an adequate operational definition of successful unlearning completeness.
    Standard in FU literature and used throughout §4; no certified indistinguishability is claimed.
  • standard math Standard SGD/ResNet-18 training dynamics and Dirichlet non-IID partitions are representative enough for the reported averages to generalize within vision FL benchmarks.
    Experimental setup §4.1; ordinary ML experimental assumption.
invented entities (3)
  • pFedUL three-stage pipeline (Fisher attribution + adaptive selective correction + lightweight recalibration) no independent evidence
    purpose: Operationalize layer-aware unlearning under the shared/personalized split so that completeness and personalization preservation can be traded off controllably.
    Core algorithmic contribution of the paper; no independent existence outside this work.
  • Personalization Preservation Score (PPS) no independent evidence
    purpose: Measure average ratio of post- to pre-unlearning personalized accuracy across remaining clients.
    Defined in Eq. 14; new evaluation construct for pFL unlearning.
  • Cross-client Fairness Index (CFI) no independent evidence
    purpose: Measure uniformity of unlearning impact via 1 − std of per-client preservation ratios.
    Defined in Eq. 15; new fairness-oriented metric.

reviewed 2026-07-15 · how reviews work

0 comments
Cite this review

Pith. "Pith review of pFedUL: Layer-Aware Federated Unlearning for Personalized Federated Learning." pith.science (2026). https://pith.science/paper/OXYWRURJ

@misc{pith2026260616304,
  author       = {Pith},
  title        = {Pith review of: pFedUL: Layer-Aware Federated Unlearning for Personalized Federated Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OXYWRURJ}},
  note         = {Machine review of arXiv:2606.16304}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

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.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

44 extracted references · 4 linked inside Pith

  1. [1]

    Communication-efficient learning of deep networks from decentralized data

    McMahan B, Moore E, Ramage D, Hampson S, Arcas BA. Communication-efficient learning of deep networks from decentralized data. In: Proceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS); 2017 Apr 20-22; Fort Lauderdale, FL, USA. p. 1273-1282

    2017

  2. [2]

    Advances and open problems in federated learning

    Kairouz P, McMahan HB, Avent B, Bellet A, Bennis M, Bhagoji AN, et al. Advances and open problems in federated learning. Found Trends Mach Learn. 2021;14(1-2):1-210. [CrossRef]

    2021

  3. [3]

    Federated learning: Challenges, methods, and future directions

    Li T, Sahu AK, Talwalkar A, Smith V. Federated learning: Challenges, methods, and future directions. IEEE Signal Process Mag. 2020;37(3):50-60. [CrossRef]

    2020

  4. [4]

    European Parliament, Council of the European Union. Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data. Off J Eur Union. 2016;L119:1-88

    2016

  5. [5]

    Machine unlearning

    Bourtoule L, Chandrasekaran V, Choquette-Choo CA, Jia H, Travers A, Zhang B, et al. Machine unlearning. In: Proceedings of the 2021 IEEE Symposium on Security and Privacy (S&P); 2021 May 24-27; San Francisco, CA, USA. p. 141-159. [CrossRef]

    2021

  6. [6]

    Understanding black-box predictions via influence functions

    Koh PW, Liang P. Understanding black-box predictions via influence functions. In: Proceedings of the 34th International Conference on Machine Learning (ICML); 2017 Aug 6-11; Sydney, Australia. p. 1885-1894

    2017

  7. [7]

    Machine unlearning: A survey

    Xu H, Zhu T, Zhang L, Zhou W, Yu PS. Machine unlearning: A survey. ACM Comput Surv. 2023;56(1):1-36. [CrossRef] Comput Mater Contin. 2026;0(0):0 27

    2023

  8. [8]

    Federated unlearning: A survey on methods, design guidelines, and evaluation metrics

    Romandini N, Mora A, Mazzocca C, Montanari R, Bellavista P. Federated unlearning: A survey on methods, design guidelines, and evaluation metrics. IEEE Trans Neural Netw Learn Syst. 2024;35(12):17513-17530. [CrossRef]

    2024

  9. [9]

    FedEraser: Enabling efficient client-level data removal from federated learning

    Liu G, Ma X, Yang Y, Wang C, Liu J. FedEraser: Enabling efficient client-level data removal from federated learning. In: Proceedings of the 2021 IEEE/ACM 29th International Symposium on Quality of Service (IWQoS); 2021 Jun 25-28; Tokyo, Japan. p. 1-10. [CrossRef]

    2021

  10. [10]

    Asynchronous federated unlearning

    Su L, Li J. Asynchronous federated unlearning. In: Proceedings of the IEEE Conference on Computer Communications (INFOCOM); 2023 May 17-20; New York, NY, USA. p. 1-10. [CrossRef]

    2023

  11. [11]

    FedRecovery: Differentially private machine unlearning for federated learning frameworks

    Zhang L, Zhu T, Calandrini P, Yin J. FedRecovery: Differentially private machine unlearning for federated learning frameworks. IEEE Trans Inf Forensics Secur. 2023;18:4732-4746. [CrossRef]

    2023

  12. [12]

    Unlearning during learning: An efficient federated machine unlearning method

    Gu H, Zhu G, Zhang J, Zhao X, Han Y, Fan L, et al. Unlearning during learning: An efficient federated machine unlearning method. In: Proceedings of the 33rd International Joint Conference on Artificial Intelligence (IJCAI); 2024 Aug 3-9; Jeju, South Korea. p. 4035-4043. [CrossRef]

    2024

  13. [13]

    Federated unlearning with gradient descent and conflict mitigation

    Pan Z, Wang Z, Li C, Zheng K, Wang B, Tang X, et al. Federated unlearning with gradient descent and conflict mitigation. In: Proceedings of the 39th AAAI Conference on Artificial Intelligence; 2025 Feb 25-Mar 4; Philadelphia, PA, USA. p. 19804-19812. [CrossRef]

    2025

  14. [14]

    Federated unlearning via class-discriminative pruning

    Wang J, Guo S, Xie X, Qi H. Federated unlearning via class-discriminative pruning. In: Proceedings of the ACM Web Conference (WWW); 2022 Apr 25-29; Lyon, France. p. 622-632. [CrossRef]

    2022

  15. [15]

    Federated unlearning: How to efficiently erase a client in FL? arXiv preprint

    Halimi A, Kadhe S, Rawat A, Baracaldo N. Federated unlearning: How to efficiently erase a client in FL? arXiv preprint. 2022;arXiv:2207.05521. [CrossRef]

    Pith/arXiv arXiv 2022

  16. [16]

    Ferrari: Federated feature unlearning via optimizing feature sensitivity

    Gu Y, Bai Y, Lyu L. Ferrari: Federated feature unlearning via optimizing feature sensitivity. In: Proceedings of the 38th Conference on Neural Information Processing Systems (NeurIPS); 2024 Dec 10-15; Vancouver, Canada. p. 1-12. [CrossRef]

    2024

  17. [17]

    A survey on federated unlearning: Challenges and opportunities

    Jeong G, Ma K. A survey on federated unlearning: Challenges and opportunities. IEEE Trans Big Data. 2026. [CrossRef]

    2026

  18. [18]

    VeriFi: Towards verifiable federated unlearning

    Gao Y, Liu Y, Xiong H, Han Z. VeriFi: Towards verifiable federated unlearning. IEEE Trans Dependable Secure Comput. 2024;21(5):4236-4249. [CrossRef]

    2024

  19. [19]

    Federated optimization in heterogeneous networks

    Li T, Sahu AK, Zaheer M, Sanjabi M, Talwalkar A, Smith V. Federated optimization in heterogeneous networks. In: Proceedings of Machine Learning and Systems (MLSys); 2020 Mar 2-4; Austin, TX, USA. p. 429-450

    2020

  20. [20]

    Towards personalized federated learning

    Tan AZ, Yu H, Cui L, Yang Q. Towards personalized federated learning. IEEE Trans Neural Netw Learn Syst. 2023;34(12):9587-9603. [CrossRef]

    2023

  21. [21]

    Federated learning with personalization layers

    Arivazhagan MG, Aggarwal V, Singh AK, Choudhary S. Federated learning with personalization layers. arXiv preprint. 2019;arXiv:1912.00818

    Pith/arXiv arXiv 2019

  22. [22]

    Exploiting shared representations for personalized federated learning

    Collins L, Hassani H, Mokhtari A, Shakkottai S. Exploiting shared representations for personalized federated learning. In: Proceedings of the 38th International Conference on Machine Learning (ICML); 2021 Jul 18-24; Virtual. p. 2089-2099

    2021

  23. [23]

    Ditto: Fair and robust federated learning through personalization

    Li T, Hu S, Beirami A, Smith V. Ditto: Fair and robust federated learning through personalization. In: Proceedings of the 38th International Conference on Machine Learning (ICML); 2021 Jul 18-24; Virtual. p. 6357-6368

    2021

  24. [24]

    FedBN: Federated learning on non-IID features via local batch normalization

    Li X, Jiang M, Zhang X, Kamp M, Deschle T. FedBN: Federated learning on non-IID features via local batch normalization. In: Proceedings of the 9th International Conference on Learning Representations (ICLR); 2021 May 3-7; Virtual. p. 1-16

    2021

  25. [25]

    FedALA: Adaptive local aggregation for personalized federated learning

    Zhang J, Hua Y, Wang H, Song T, Xue Z, Ma R, et al. FedALA: Adaptive local aggregation for personalized federated learning. In: Proceedings of the 37th AAAI Conference on Artificial Intelligence; 2023 Feb 7-14; Washington, DC, USA. p. 11237-11244. [CrossRef]

    2023

  26. [26]

    Zero-shot federated unlearning via transforming from data-dependent to personalized model-centric

    Wu W, Liang H, Yuan J, Jiang J, Wang KY, Hu C, et al. Zero-shot federated unlearning via transforming from data-dependent to personalized model-centric. In: Proceedings of the 34th International Joint Conference on Artificial Intelligence (IJCAI); 2025 Aug 16-22; Montreal, Canada. p. 6588-6596. [CrossRef]

    2025

  27. [27]

    Mimir: Data-free federated unlearning through client-specific prompt generation for personalized models

    Wu W, Liang H, Tu T, Jiang J, Hu C, Cheng D. Mimir: Data-free federated unlearning through client-specific prompt generation for personalized models. IEEE Trans Mob Comput. 2025;24(10):10537-10556. [CrossRef]

    2025

  28. [28]

    Unlearning through knowledge overwriting: Reversible federated unlearning via selective sparse adapter

    Zhong Z, Bao W, Wang J, Zhang S, Zhou J, Lyu L, et al. Unlearning through knowledge overwriting: Reversible federated unlearning via selective sparse adapter. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); 2025 Jun 10-17; Nashville, TN, USA. p. 30661-30670. [CrossRef] 28 Comput Mater Contin. 2026;0(0):0

    2025

  29. [29]

    A survey of machine unlearning

    Nguyen TT, Huynh TT, Nguyen PL, Liew AWC, Yin H, Nguyen QVH. A survey of machine unlearning. ACM Comput Surv. 2024;56(12):1-37. [CrossRef]

    2024

  30. [30]

    Remember what you want to forget: Algorithms for machine unlearning

    Sekhari A, Acharya J, Kamath G, Suresh AT. Remember what you want to forget: Algorithms for machine unlearning. In: Proceedings of the 35th Conference on Neural Information Processing Systems (NeurIPS); 2021 Dec 6-14; Virtual. p. 18075-18086

    2021

  31. [31]

    Unrolling SGD: Understanding factors of influence in machine unlearning

    Thudi A, Deza G, Chandrasekaran V, Papernot N. Unrolling SGD: Understanding factors of influence in machine unlearning. In: Proceedings of the IEEE Conference on Secure and Trustworthy Machine Learning (SaTML); 2022; Virtual. p. 1-13. [CrossRef]

    2022

  32. [32]

    Overcoming catastrophic forgetting in neural networks

    Kirkpatrick J, Pascanu R, Rabinowitz N, Veness J, Desjardins G, Rusu AA, et al. Overcoming catastrophic forgetting in neural networks. Proc Natl Acad Sci. 2017;114(13):3521-3526. [CrossRef]

    2017

  33. [33]

    The right to be forgotten in federated learning: An efficient realization with rapid retraining

    Liu Y, Xu L, Yuan X, Wang C, Li B. The right to be forgotten in federated learning: An efficient realization with rapid retraining. In: Proceedings of the IEEE Conference on Computer Communications (INFOCOM); 2022 May 2-5; London, United Kingdom. p. 1749-1758. [CrossRef]

    2022

  34. [34]

    Machine unlearning of federated clusters

    Pan C, Sima J, Prakash S, Rana V, Milenkovic O. Machine unlearning of federated clusters. In: Proceedings of the 11th International Conference on Learning Representations (ICLR); 2023 May 1-5; Kigali, Rwanda. p. 1-27

    2023

  35. [35]

    NoT: Federated unlearning via weight negation

    Khalil A, Souici S, Boutiche M, Erbad A, Mohamed A. NoT: Federated unlearning via weight negation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); 2025 Jun 10-17; Nashville, TN, USA. p. 30661-30670. [CrossRef]

    2025

  36. [36]

    Fast federated machine unlearning with nonlinear functional theory

    Che T, Zhou Z, Lyu L, Xia J, Chen J, Yan B, et al. Fast federated machine unlearning with nonlinear functional theory. In: Proceedings of the 40th International Conference on Machine Learning (ICML); 2023 Jul 23-29; Honolulu, HI, USA. p. 4241-4268

    2023

  37. [37]

    Federated unlearning for on-device recommendation

    Yuan W, Yin H, Wu F, Zhang S, He T, Wang H. Federated unlearning for on-device recommendation. In: Proceedings of the 16th ACM International Conference on Web Search and Data Mining (WSDM); 2023 Feb 27-Mar 3; Singapore. p. 393-401. [CrossRef]

    2023

  38. [38]

    Sequential informed federated unlearning: Efficient and provable client unlearning in federated optimization

    Fraboni Y, Vidal R, Kameni L, Lorenzi M. Sequential informed federated unlearning: Efficient and provable client unlearning in federated optimization. In: Proceedings of the 27th International Conference on Artificial Intelligence and Statistics (AISTATS); 2024 May 2-4; Valencia, Spain. p. 1-12

    2024

  39. [39]

    Federated learning on non-IID data silos: An experimental study

    Li Q, Diao Y, Chen Q, He B. Federated learning on non-IID data silos: An experimental study. In: Proceedings of the 38th IEEE International Conference on Data Engineering (ICDE); 2022 May 9-12; Kuala Lumpur, Malaysia. p. 965-978. [CrossRef]

    2022

  40. [40]

    On bridging generic and personalized federated learning for image classification

    Chen HY, Chao WL. On bridging generic and personalized federated learning for image classification. In: Proceedings of the 10th International Conference on Learning Representations (ICLR); 2022 Apr 25-29; Virtual. p. 1-17

    2022

  41. [41]

    Personalized federated learning with Moreau envelopes

    T Dinh C, Tran NH, Nguyen TD. Personalized federated learning with Moreau envelopes. In: Proceedings of the 34th Conference on Neural Information Processing Systems (NeurIPS); 2020 Dec 6-12; Virtual. p. 21394-21405

    2020

  42. [42]

    Membership inference attacks against machine learning models

    Shokri R, Stronati M, Song C, Shmatikov V. Membership inference attacks against machine learning models. In: Proceedings of the 2017 IEEE Symposium on Security and Privacy (S&P); 2017 May 22-26; San Jose, CA, USA. p. 3-18. [CrossRef]

    2017

  43. [43]

    LEAF: A benchmark for federated settings

    Caldas S, Duddu SMK, Wu P, Li T, Konecny J, McMahan HB, et al. LEAF: A benchmark for federated settings. arXiv preprint. 2019;arXiv:1812.01097

    Pith/arXiv arXiv 2019

  44. [44]

    Measuring the effects of non-independently and identically distributed data on federated learning

    Hsu TMH, Qi H, Brown M. Measuring the effects of non-independently and identically distributed data on federated learning. arXiv preprint. 2022;arXiv:2009.09890

    Pith/arXiv arXiv 2022

This paper was first reviewed by grok-4.5 on July 15, 2026.