Pith. sign in

REVIEW 4 major objections 7 minor 43 references

Vertical Federated Continual Learning via Evolving Prototype Knowledge

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper proposes V-LETO, a method that lets vertical federated learning models add new classes and new features over time without catastrophic forgetting, by evolving class prototypes and freezing important local parameters.

desk verdict The FIL half is coherent and the problem framing is novel, but the CIL prototype-evolution equation is undefined as written, so the headline CIL claims are not yet supported. read the letter →

arxiv 2502.09152 v2 pith:7TADTNEK submitted 2025-02-13 cs.LG cs.NE

Shuo Wang , Keke Gai , Jing Yu , Liehuang Zhu , Qi Wu This is my paper · ORCID
classification cs.LGcs.NE
keywords verticalfederatedlearningcontinualcatastrophicforgettingclassincrementalfeatureprototypeknowledgeFisherinformationmatrix
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Vertical federated learning (VFL) joins parties that hold different features of the same samples while only one party sees labels; this paper asks whether such a system can keep learning when new classes arrive and when new feature columns are added, without erasing what it already learned. The answer it proposes, V-LETO, stores compact per-class prototypes built from aggregated embeddings and evolves those prototypes across tasks, while passive parties freeze the local parameters that mattered most to previous tasks. If the central claim is right, V-LETO would be the first VFL method to handle class-incremental and feature-incremental learning together, and it would let finance and healthcare deployments add new labels and new data sources over time without retraining from scratch or sharing raw data. The paper reports that V-LETO beats the strongest adapted baseline by 10.39 percentage points on class-incremental average accuracy and by 35.15 percentage points on feature-incremental average accuracy across four benchmark datasets. The practical stakes are that a deployed VFL model could keep improving as the world changes instead of being rebuilt whenever a new class or column appears.

What carries the argument

The load-bearing object is the evolving global prototype list, a set of per-class mean embeddings stored at the active party. It compresses previous-task knowledge without raw data, generates synthetic old-class vectors for class-incremental learning, and supplies weighted prior-feature prototypes for feature-incremental learning. The second piece is the local model optimizer: a per-parameter Fisher Information estimate, computed from squared gradients, selects which passive-party parameters to freeze so that updates for the current task do not erase the parameters most significant to earlier tasks.

What would settle it

Run V-LETO on a class-incremental sequence; after the current task, take held-out old-class samples, compute their true mean global embedding, and compare its distance to the pseudo-prototype from Eq. (2) and to the unmodified stored prototype. If the pseudo-prototype is not closer to the true drifted embedding, or if using the unmodified prototype preserves old-class accuracy just as well, then the class-incremental evolution rule is not doing the work attributed to it.

Watch

Extended reading notes

Core claim

V-LETO claims that catastrophic forgetting in vertical federated learning can be prevented with an evolving global prototype list, without replaying raw data from earlier tasks. On each task, the active party sums passive-party local embeddings into global embeddings and averages them per class to make prototypes; the Class-Incremental branch synthesizes pseudo-prototypes for old classes by adding a cosine-similarity-scaled displacement to the stored prototype, and the Feature-Incremental branch blends the current class prototype with the stored one. The global model then trains on the real embeddings plus prototype-based losses for old and new knowledge, and each passive party computes a Fisher Information estimate per parameter and freezes the parameters above a threshold so local updates do not overwrite prior-task knowledge. The reported result is that class accuracy no longer collapses as tasks grow, and feature accuracy rises monotonically from task 1 to task 4 across the four datasets, with the largest margin over baselines in the feature-incremental setting.

Load-bearing premise

The load-bearing premise is that adding a cosine-similarity-scaled displacement to a stored old-class prototype produces a trustworthy stand-in for where that class's embedding actually moved after new classes were learned; nothing in the paper justifies the direction or size of that displacement.

Editorial extensions

If this is right

  • A VFL deployment can receive a stream of new classes without needing old data or old labels at the passive parties, because the prototypes carry the prior knowledge.
  • A VFL deployment can receive new feature columns without retraining on the full historical feature set, because the feature-incremental branch mixes current and stored prototypes.
  • Passive parties do not need old gradients or old data to protect prior knowledge; freezing Fisher-selected parameters is enough.
  • Because only prototypes and embeddings are shared, the method avoids the raw-data replay that would otherwise violate the privacy assumptions of VFL.
  • If the reported gains hold, the method gives an upgrade path for existing VFL systems rather than a redesign: add the prototype modules and the freezing rule to the current pipeline.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The class-incremental pseudo-prototype rule is the most fragile component: adding a cosine-similarity-scaled shift assumes that similarity between old and new prototypes points in the direction of the old class's drift, which will fail for class orders where new classes are unlike old ones; swapping the task order should change the reported margins if this is the active mechanism.
  • The feature-incremental update is essentially a weighted running average of prototypes; comparing V-LETO against a simple exponential moving average with the same beta would isolate how much of the 35.15 percentage point gain comes from prototype weighting versus from the frozen local parameters.
  • A natural extension is to add differential privacy to the shared global embeddings; privacy leakage through the prototype list is not analyzed in this work, so that boundary remains open.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The manuscript proposes V-LETO, a vertical federated continual learning method that combines three modules: prototype generation from global embeddings, evolving prototype knowledge for class-incremental (CIL) and feature-incremental (FIL) tasks, and Fisher-information-based local parameter freezing. The authors claim that V-LETO is the first VFL method to handle both CIL and FIL, and report experiments on FMNIST, CIFAR10, CINIC10, and NEWS20 showing large accuracy gains over adapted baselines, together with ablations, hyperparameter analyses, and t-SNE visualizations.

Significance. If the results hold, the paper addresses an underexplored and practically important setting, and the FIL results in Table 2 are striking: V-LETO shows continuous accuracy growth as features are added, unlike all baselines. The method is also clearly described at a high level, and the authors provide an anonymized code link, which is a positive step for reproducibility. However, the central CIL mechanism is ill-defined as written, and the headline improvement numbers cannot be traced to the tables, so the current version does not yet provide a sound empirical or methodological basis for the paper's strongest claims.

major comments (4)
  1. [Prototype Evolving, Eq. (2)] Equation (2) computes a sum over c ∈ C_t (current-task classes) of dis(µ^{t-1}_c, µ^t_c), but the paper's CIL protocol specifies that each task introduces new, disjoint classes, so for any c ∈ C_t there is no prototype µ^{t-1}_c from the previous task. The text immediately before Eq. (2) also states p ∉ C_t, confirming that the sum runs over classes not seen before. Consequently, the arguments to dis are undefined, and the CI-PE module cannot be implemented as described. This is a load-bearing flaw because the reported CIL improvements (Table 1, abstract) are explicitly attributed to this module.
  2. [Prototype Evolving, Eq. (3)] Equation (3) defines dis as cosine similarity, yet Eq. (2) uses it as an additive displacement to the stored prototype µg_p. Cosine similarity is a similarity score in [-1,1], not a distance, and the paper gives no derivation or motivation for why adding the average cosine similarity (which can be negative) between previous and current prototypes produces a valid pseudo-prototype for an old class. Even if the indexing problem in Eq. (2) were corrected, the direction, magnitude, and sign handling of the added vector are not justified.
  3. [Abstract and Comparison with State-of-the-Art Methods] The abstract's claim of improvements of 10.39% (CIL) and 35.15% (FIL) over the state-of-the-art does not clearly match Tables 1 and 2. For CIL, the average-accuracy margins over the best baseline across the four datasets are approximately 2.2, 12.7, 8.0, and 5.0 percentage points (FMNIST, CIFAR10, CINIC10, NEWS20), not 10.39. For FIL, 35.15% appears to be the relative improvement on a single dataset/task (CIFAR10 T4: 88.34 vs. 65.36), but the text does not state this. The paper must specify which baseline, dataset, task, and absolute-vs-relative convention are used for every headline number.
  4. [Tables 1, 2, 3, 5, and 6] All experimental results are reported as point estimates without standard deviations, error bars, or the number of random seeds. Given that several improvements are modest (e.g., FMNIST CIL average 76.14 vs. 73.97, NEWS20 CIL 56.60 vs. 51.59), and the hyperparameters were tuned on FMNIST before being applied to other datasets, the statistical significance of the claimed gains is unclear. The authors should report mean ± std over at least three independent runs and, where appropriate, a significance test.
minor comments (7)
  1. [Prototype Evolving (text and Eq. (3))] The text repeatedly refers to a "distance" between prototypes, but Eq. (3) defines cosine similarity; either replace the term with a proper distance (e.g., 1 − cosine similarity) or explain why cosine similarity itself is the intended quantity.
  2. [Model Optimization] The subsection title "Gobal Model Optimization" contains a typo; it should read "Global Model Optimization."
  3. [Implementation Details] The hyperparameter values are incompletely reported: Eq. (2) introduces γ and Eq. (5) introduces λCE, but the implementation details paragraph gives values only for β, λF, λA, k0, and α, with no default for γ or λCE; Table 4 varies λCE but does not state the fixed values of λF and γ used in those runs.
  4. [Eq. (7) and surrounding text] The sentence "δ = k0 + α log(t+1) is a hyperparameter" is imprecise; δ is a function of the hyperparameters k0 and α, not itself a directly set hyperparameter, and the paper should clarify which quantity is actually tuned.
  5. [Table 3 caption] The evaluation protocol for the ablation table is hard to follow; the caption should explain how "T12" or "T123" accuracies are computed (e.g., joint evaluation on all classes from tasks 1 and 2) and what each row/column block represents.
  6. [Model Optimization (auxiliary losses)] The auxiliary losses LA and LF in Eq. (5) use prototypes generated from the model's own embeddings through the PG module, so they function as self-consistency regularization rather than externally grounded knowledge transfer; the paper should discuss this characterization explicitly, as the phrase "transfer of knowledge from previous tasks" overstates the mechanism.
  7. [Code availability] The abstract states that code is available through an anonymous link; if possible, provide a permanent repository URL or state that a cleaned version will be released in a public repository upon acceptance.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: the reported gains are empirical comparisons, and the evolving-prototype losses are self-distillation targets rather than derived predictions.

full rationale

The manuscript supports its central claim with held-out test accuracy across four datasets against adapted baselines, so the headline CIL/FIL improvements are not defined by construction from the method's own inputs. The prototype generation step builds class prototypes from current global embeddings, and the CI-PE/FI-PE losses L_A and L_F use those evolved prototypes as training targets; this is a self-referential optimization objective, but the benchmark results are evaluated on real data and labels, so the comparison does not reduce to a fitted parameter or a tautology. The CI-PE equation is a specification concern: Eq. (2)-(3) call for dis(µ^{t-1}_c, µ^t_c) for c in the current task's class set, yet the paper's disjoint-class CIL protocol means those previous-task prototypes do not exist; that is an undefined-input reproducibility issue, not an input/output equivalence. The self-citations (Wang et al. 2025a/b/c) appear only as general VFL background and do not carry the central argument. Hyperparameters tuned on FMNIST and applied to other sets are mild in-domain selection, not circular derivation. Overall, no step in the derivation chain is forced by definition or by a self-citation loop.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The method relies on standard continual-learning assumptions (unavailability of old data, FIM approximating parameter importance) plus several ad-hoc heuristics specific to this paper: the prototype evolution rule that adds cosine similarity, the delta = k0 + alpha log(t+1) threshold schedule, and the choice of prototype aggregation weights. The free parameters are hyperparameters tuned on FMNIST.

free parameters (7)
  • gamma = not reported
    Weighting factor for the knowledge distance in Eq. 2 for class-incremental prototype evolution; not listed in the implementation details.
  • beta = 0.5
    Weighting factor in Eq. 4 for aggregating current and previous feature prototypes.
  • lambda_CE = 0.5
    Weighting of the current-task cross-entropy loss in Eq. 5; Table 4 shows 0.5 gives the best T1234 accuracy.
  • lambda_A = 0.5
    Weighting of the class-prototype loss in Eq. 5; set to 0.5.
  • lambda_F = 0.5
    Weighting of the feature-prototype loss in Eq. 5; set to 0.5.
  • k0 = 15
    Base offset in the threshold schedule delta = k0 + alpha log(t+1) in Eq. 7; chosen via Table 6.
  • alpha = 3
    Multiplicative term in the threshold schedule in Eq. 7; Table 6 shows alpha=3 paired with k0=15 gives the best result, although the text says alpha=5.
assumptions (5)
  • domain assumption Global embedding is the sum of local embeddings.
    Used in Prototype Generation: E_i^t = sum of E_ik^t over passive parties. This is a standard VFL aggregation choice.
  • ad hoc to paper Cosine similarity can serve as a distance measure for prototype evolution.
    Eqs. 2 and 3 add a scaled cosine similarity to prototypes, treating similarity as a displacement distance without justification.
  • domain assumption FIM on current-task gradients identifies parameters important for previous tasks.
    Standard EWC-style assumption used in Local Model Optimization, Eq. 6.
  • ad hoc to paper delta = k0 + alpha log(t+1) with larger delta putting more emphasis on previous tasks.
    Eq. 7 introduces a heuristic threshold schedule with no derivation.
  • domain assumption Previous task data is unavailable and cannot be replayed.
    The problem definition states data from previous tasks becomes unavailable once a new task arrives; replay is not used.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Vertical Federated Continual Learning via Evolving Prototype Knowledge." pith.science (2026). https://pith.science/paper/7TADTNEK

@misc{pith2026250209152,
  author       = {Pith},
  title        = {Pith review of: Vertical Federated Continual Learning via Evolving Prototype Knowledge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7TADTNEK}},
  note         = {Machine review of arXiv:2502.09152}
}
read the original abstract

Vertical Federated Learning (VFL) has garnered significant attention as a privacy-preserving machine learning framework for sample-aligned feature federation. However, traditional VFL approaches do not address the challenges of class and feature continual learning, resulting in catastrophic forgetting of knowledge from previous tasks. To address the above challenge, we propose a novel vertical federated continual learning method, named Vertical Federated Continual Learning via Evolving Prototype Knowledge (V-LETO), which primarily facilitates the transfer of knowledge from previous tasks through the evolution of prototypes. Specifically, we propose an evolving prototype knowledge method, enabling the global model to retain both previous and current task knowledge. Furthermore, we introduce a model optimization technique that mitigates the forgetting of previous task knowledge by restricting updates to specific parameters of the local model, thereby enhancing overall performance. Extensive experiments conducted in both CIL and FIL settings demonstrate that our method, V-LETO, outperforms the other state-of-the-art methods. For example, our method outperforms the state-of-the-art method by 10.39% and 35.15% for CIL and FIL tasks, respectively. Our code is available at https://anonymous.4open.science/r/V-LETO-0108/README.md.

Figures

Figures reproduced from arXiv: 2502.09152 by the authors.

Figure 1
Figure 1. An illustration of CIL and FIL in VFL. Task 2 in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The framework overview of V-LETO. V-LETO consists of three modules: [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The impact of hyperparameters (λCE, λF ) on FIL on the FMNIST dataset. and can hardly use features from previous tasks, while our method has been evidenced that it can effectively utilize fea￾tures from previous tasks to optimize model parameters and improve performance. Ablation Study For CIL in V-LETO, [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: t-SNE visualization of the global embedding and prototypes produced by CIL (a)(b)(c)(d) and FIL (e)(f)(g)(h) in V-LETO on the FMNIST dataset. Colored circles represent global embeddings, and black squares denote class prototypes. Hyper-parameter Analysis We evaluated t…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 39 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    S.; Rady, H.; and Moustafa, M

    Abouelnaga, Y.; Ali, O. S.; Rady, H.; and Moustafa, M. 2016. Cifar-10 : Knn-based ensemble of classifiers. In 2016 International Conference on Computational Science and Computational Intelligence, 1192--1195. Las Vegas, NV, USA

    work page 2016

  4. [4]

    T.; and Ma, R

    Cai, T. T.; and Ma, R. 2022. Theoretical foundations of t-sne for visualizing high-dimensional clustered data. Journal of Machine Learning Research, 23(301): 1--54

    work page 2022

  5. [5]

    E.; Lema, D.; Iglesias, R.; Regueiro, C

    Casado, F. E.; Lema, D.; Iglesias, R.; Regueiro, C. V.; and Barro, S. 2023. Ensemble and continual federated learning for classification tasks. Machine Learning, 112(9): 3413--3453

    work page 2023

  6. [6]

    J.; Das, A.; Wang, S.; and Patterson, S

    Castiglia, T. J.; Das, A.; Wang, S.; and Patterson, S. 2022. Compressed-VFL : Communication-efficient learning with vertically partitioned data. In International Conference on Machine Learning, 2738--2766. Baltimore, Maryland, USA

    work page 2022

  7. [7]

    a., Boyan Wei

    Feng Qiang, e. a., Boyan Wei. 2022. White Paper on the Application of Federated Learning Technology in Finance. Http://www.hbbill.com/uploadFiles/-16/548/058/54/

    work page 2022

  8. [8]

    Gao, X.; Yang, X.; Yu, H.; Kang, Y.; and Li, T. 2024. FedProK: Trustworthy Federated Class-Incremental Learning via Prototypical Feature Knowledge Transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 4205--4214. Seattle, WA, USA

    work page 2024

Show all 43 references
  1. [9]

    Hou, C.; Gu, S.; Xu, C.; and Qian, Y. 2023. Incremental Learning for Simultaneous Augmentation of Feature and Class. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12): 14789--14806

    2023

  2. [10]

    Hu, C.; Chen, Y.; Peng, X.; Yu, H.; Gao, C.; and Hu, L. 2019. A Novel Feature Incremental Learning Method for Sensor-Based Activity Recognition. IEEE Transactions on Knowledge and Data Engineering, 31(6): 1038--1050

    2019

  3. [11]

    Krawczyk, A.; and Gepperth, A. 2024. An analysis of best-practice strategies for replay and rehearsal in continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 4196--4204. Seattle,WA, USA

    2024

  4. [12]

    Lang, K. 1995. Newsweeder: Learning to filter netnews. In Machine learning proceedings 1995, 331--339. Elsevier

    1995

  5. [13]

    M.; Le Borgne, Y.-A.; Obl \'e , F.; and Bontempi, G

    Lebichot, B.; Siblini, W.; Paldino, G. M.; Le Borgne, Y.-A.; Obl \'e , F.; and Bontempi, G. 2024. Assessment of catastrophic forgetting in continual credit card fraud detection. Expert Systems with Applications, 249(99): 123445

    2024

  6. [14]

    Li, S.; Su, T.; Zhang, X.; and Wang, Z. 2024 a . Continual learning with knowledge distillation: A survey. IEEE Transactions on Neural Networks and Learning Systems, 36(6): 9798 -- 9818

    2024

  7. [15]

    Li, Y.; Li, Q.; Wang, H.; Li, R.; Zhong, W.; and Zhang, G. 2024 b . Towards Efficient Replay in Federated Incremental Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 12820--12829. Seattle, WA, USA

    2024

  8. [16]

    K.; Huang, H.; and Zheng, Z

    Liao, T.; Fu, L.; Zhang, L.; Yang, L.; Chen, C.; Ng, M. K.; Huang, H.; and Zheng, Z. 2025. Privacy-preserving vertical federated learning with tensor decomposition for data missing features. IEEE Transactions on Information Forensics and Security, 20: 3445 -- 3460

    2025

  9. [17]

    Liu, Y.; Kang, Y.; Zou, T.; Pu, Y.; He, Y.; Ye, X.; Ouyang, Y.; Zhang, Y.-Q.; and Yang, Q. 2024. Vertical federated learning: Concepts, advances, and challenges. IEEE Transactions on Knowledge and Data Engineering, 36(7): 3615 -- 3634

    2024

  10. [18]

    Liu, Z.; Sun, H.; Song, J.; Zhang, B.; Yan, Y.; Qiu, B.; Jiang, L.; and Li, J. 2023. Vertical Federated Learning Architecture for Power Company and Financial Company and Electricity Pricing Model Considering User Credit Evaluation. In 2023 3rd International Conference on Consu...

    2023

  11. [19]

    Luo, K.; Li, X.; Lan, Y.; and Gao, M. 2023. Gradma: A gradient-memory-based accelerated federated learning with alleviated catastrophic forgetting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 3708--3717. Seattle, WA, USA

    2023

  12. [20]

    Luo, M.; Chen, F.; Hu, D.; Zhang, Y.; Liang, J.; and Feng, J. 2021 a . No Fear of Heterogeneity: Classifier Calibration for Federated Learning with Non-IID Data. In Advances in Neural Information Processing Systems, 5972--5984. Virtual

    2021

  13. [21]

    Luo, X.; Wu, Y.; Xiao, X.; and Ooi, B. C. 2021 b . Feature inference attack on model predictions in vertical federated learning. In 2021 IEEE 37th International Conference on Data Engineering, 181--192. Chania, Greece

    2021

  14. [22]

    Ma, Y.; Xie, Z.; Wang, J.; Chen, K.; and Shou, L. 2022. Continual Federated Learning Based on Knowledge Distillation. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, 2182--2188. Vienna, Austria

    2022

  15. [23]

    Ni, H.; Gu, S.; Fan, R.; and Hou, C. 2024. Feature incremental learning with causality. Pattern Recognition, 146(99): 110033

    2024

  16. [24]

    Qiao, F.; and Mahdavi, M. 2024. Learn more, but bother less: parameter efficient continual learning. Advances in Neural Information Processing Systems, 37: 97476--97498

    2024

  17. [25]

    Qiu, P.; Pu, Y.; Liu, Y.; Liu, W.; Yue, Y.; Zhu, X.; Li, L.; Li, J.; and Ji, S. 2024. Integer Is Enough: When Vertical Federated Learning Meets Rounding. In Proceedings of the AAAI Conference on Artificial Intelligence, 14704--14712. Vancouver, Canada

    2024

  18. [26]

    J.; Papadopoulos, P.; Titcombe, T.; Ismail, A.; Cebere, T.; Sandmann, R.; Roehm, R.; and Hoeh, M

    Romanini, D.; Hall, A. J.; Papadopoulos, P.; Titcombe, T.; Ismail, A.; Cebere, T.; Sandmann, R.; Roehm, R.; and Hoeh, M. A. 2021. Pyvertical: A vertical federated learning framework for multi-headed splitnn. arXiv preprint arXiv:2104.00489, PP(99): 1--9

    2021 arXiv

  19. [27]

    K.; and Das, A

    Sakib, S. K.; and Das, A. B. 2024. Explainable Vertical Federated Learning for Healthcare: Ensuring Privacy and Optimal Accuracy. In 2024 IEEE International Conference on Big Data, 5068--5077. Washington, DC, USA

    2024

  20. [28]

    Shenaj, D.; Toldo, M.; Rigon, A.; and Zanuttigh, P. 2023. Asynchronous federated continual learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5055--5063. Vancouver, BC, Canada

    2023

  21. [29]

    Wang, G.; Gu, B.; Zhang, Q.; Li, X.; Wang, B.; and Ling, C. X. 2023. A unified solution for privacy and communication efficiency in vertical federated learning. In Advances in Neural Information Processing Systems, 13480--13491. New Orleans, LA, USA

    2023

  22. [30]

    Wang, L.; Zhang, Z.; Huang, M.; Gai, K.; Wang, J.; and Shen, Y. 2025 a . RoPA: Robust Privacy-Preserving Forward Aggregation for Split Vertical Federated Learning. IEEE Transactions on Network and Service Management, 1

    2025

  23. [31]

    Wang, Q.; Liu, B.; and Li, Y. 2024. Traceable Federated Continual Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 12872--12881. Seattle, WA, USA

    2024

  24. [32]

    Wang, S.; Gai, K.; Yu, J.; Zhang, Z.; and Zhu, L. 2025 b . Pravfed: Practical heterogeneous vertical federated learning via representation learning. IEEE Transactions on Information Forensics and Security, 20(1): 2693 -- 2705

    2025

  25. [33]

    Wang, S.; Gai, K.; Yu, J.; Zhang, Z.; and Zhu, L. 2025 c . PraVFed: Practical Heterogeneous Vertical Federated Learning via Representation Learning. IEEE Transactions on Information Forensics and Security, PP(99): 1

    2025

  26. [34]

    Xiao, H.; Rasul, K.; and Vollgraf, R. 2017. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, PP(99): 1

    2017 arXiv

  27. [35]

    Yang, X.; Huang, W.; and Ye, M. 2023. Dynamic Personalized Federated Learning with Adaptive Differential Privacy. In Advances in Neural Information Processing Systems, 72181--72192. New Orleans, LA, USA

    2023

  28. [36]

    Yang, X.; Yu, H.; Gao, X.; Wang, H.; Zhang, J.; and Li, T. 2024. Federated continual learning via knowledge fusion: A survey. IEEE Transactions on Knowledge and Data Engineering, 38(8): 3832--3850

    2024

  29. [37]

    Ye, M.; Shen, W.; Du, B.; Snezhko, E.; Kovalev, V.; and Yuen, P. C. 2025. Vertical federated learning for effectiveness, security, applicability: A survey. ACM Computing Surveys, 57(9): 1--32

    2025

  30. [38]

    Yoon, J.; Jeong, W.; Lee, G.; Yang, E.; and Hwang, S. J. 2021. Federated continual learning with weighted inter-client transfer. In International Conference on Machine Learning, 12073--12086. Virtual Event

    2021

  31. [39]

    Yu, H.; Yang, X.; Gao, X.; Feng, Y.; Wang, H.; Kang, Y.; and Li, T. 2024. Overcoming Spatial-Temporal Catastrophic Forgetting for Federated Class-Incremental Learning. In ACM Multimedia 2024, 1--9. Melbourne, Australia

    2024

  32. [40]

    Zhang, J.; Guo, S.; Qu, Z.; Zeng, D.; Wang, H.; Liu, Q.; and Zomaya, A. Y. 2022 a . Adaptive vertical federated learning on unbalanced features. IEEE Transactions on Parallel and Distributed Systems, 33(12): 4006--4018

    2022

  33. [41]

    Zhang, Z.; Guo, B.; Sun, W.; Liu, Y.; and Yu, Z. 2022 b . Cross-FCL : Toward a cross-edge federated continual learning framework in mobile edge computing systems. IEEE Transactions on Mobile Computing, 23(1): 313--326

    2022

  34. [42]

    Zhu, F.; Zhang, X.-Y.; Wang, C.; Yin, F.; and Liu, C.-L. 2021. Prototype augmentation and self-supervision for incremental learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5871--5880. Virtual

    2021

  35. [43]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence '...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.