Pith. sign in

REVIEW 4 major objections 8 minor 88 references

Proto-EVFL: Enhanced Vertical Federated Learning via Dual Prototype with Extremely Unaligned Data

T0 review · 4 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Proto-EVFL makes vertical federated learning work with scarce aligned labels by selecting unaligned samples via dual class prototypes and mixed priors, and proves convergence at rate 1/√T.

desk verdict Proto-EVFL has a solid few-shot/imbalance core, but its headline zero-shot claim cannot be produced by the algorithm as written—Eq. 15 divides by zero for an unseen class and the classifier has no output weight for it. read the letter →

arxiv 2507.22488 v1 pith:2GSK6SRS submitted 2025-07-30 cs.LG cs.AI

Wei Guo , Yiyang Duan , Zhaojun Hu , Yiqi Tong , Fuzhen Zhuang , Xiao Zhang , Jin Dong , Ruofan Wu
show 2 more authors
Tengfei Liu Yifan Sun
This is my paper
classification cs.LGcs.AI
keywords verticalfederatedlearningclassprototypeimbalancezero-shotoptimaltransportbi-leveloptimizationfeatureaggregationunaligneddata
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

Proto-EVFL is built around a simple claim: in vertical federated learning, the huge pool of samples that are not aligned across parties, and therefore have no labels, can be converted into training signal if each party learns class prototypes and uses them to pseudo-label its own unaligned data. The paper identifies two failure modes that have blocked this: intra-party class imbalance, which biases each local feature extractor toward majority classes, and inter-party class imbalance, which makes parties contribute inconsistent features to the active party's classifier. To address them, Proto-EVFL selects unaligned samples with a probabilistic dual optimal transport cost that mixes local and global class priors, and aggregates party features through an adaptive gating network. The paper proves the resulting bi-level optimization converges at rate $1/\sqrt{T}$, and reports that in a zero-shot setting with one unseen class it beats baselines by at least 6.97%. A sympathetic reader would care because, if correct, alignment scarcity stops being a hard barrier to VFL deployment.

What carries the argument

The load-bearing object is the class prototype: a per-party cluster center in latent space that stands for a class, updated on the active party using aligned representations and sent back to passive parties. The selection rule is the probabilistic dual transport cost (PDTC), which computes two conditional optimal transport costs, $L^m_{f\to\mu}$ from unaligned representations to prototypes and $L^m_{\mu\to f}$ from prototypes to representations, both using a class prior probability to reweight assignments. The mixed prior guided module sets that prior as a convex combination of the local EM-estimated prior and a global prior with a personalized weight $\gamma_m$. A gating network with per-party adaptors then weights the concatenated features before classification. Together they turn the training objective into a bi-level problem, outer classifier loss and inner extractor losses, whose convergence is analyzed through strong convexity of the regularized local loss.

What would settle it

Run Proto-EVFL on a vertical split where classes overlap rather than separate in each party's feature space, for example Fashion-MNIST quadrants with confusable classes such as pullover and coat, and compare zero-shot accuracy against the same baselines; if the at-least-6.97% margin disappears or reverses, the cluster assumption is what carries the result.

Watch

Extended reading notes

Core claim

The paper's central claim is that extremely unaligned, unlabeled data can be used in vertical federated learning without inheriting class-imbalance damage, provided each party represents classes by prototypes in its own latent space and selects which unaligned samples to trust via a dual optimal transport cost conditioned on class priors. In the forward direction (samples to prototypes) the selection cost is an entropy-minimizing assignment; in the reverse direction (prototypes to samples) it prevents any class prototype, especially a rare or unseen one, from being abandoned. A mixed prior module blends each party's estimated local class distribution with the global distribution, and an adaptive gated aggregation weights each party's adapted features before the active party's classifier. The paper states this is the first bi-level optimization formulation of VFL, proves a convergence rate of $1/\sqrt{T}$, and reports consistent improvements over baselines, including 6.97% or more over the strongest baseline in a one-unseen-class zero-shot setting.

Load-bearing premise

The load-bearing premise is that within each party's latent space, unaligned samples form separated clusters that share labels, so a prototype can stand for a class and pseudo-labels from the dual transport cost are trustworthy.

Editorial extensions

If this is right

  • Aligned sample scarcity stops being a hard blocker: each party can enlarge its extractor training set with its own unaligned, unlabeled samples, so VFL can start from as few as 200 aligned samples and still train a usable classifier.
  • Rare and unseen classes become addressable: the reverse prototype-to-sample cost keeps disappearing classes represented, which is what the paper's zero-shot gains of at least 6.97% over baselines rely on.
  • The active party can aggregate features from parties with very different local class distributions without letting any one party dominate, because the gating network reweights contributions every round.
  • The bi-level formulation gives a theoretical handle: under smoothness, bounded variance, and strong convexity of the local loss, the algorithm reaches an $\epsilon$-accurate stationary point at rate $1/\sqrt{T}$, matching the classic VFL rate.
  • Because only prototypes and intermediate representations are exchanged, gradient-based label and feature inference attacks lose their main foothold; the paper reports label-inference attack accuracy dropping to 8.89%.

Reading between the lines

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

  • Editorial inference: the dual-prototype selection rule is a general pseudo-labeling strategy, not a VFL-specific trick; it could be lifted to any semi-supervised or domain-adaptation setting with heterogeneous feature spaces and class-imbalanced pools of unlabeled data.
  • Editorial inference: the zero-shot gains are bounded by the cluster assumption; on data whose classes overlap in latent space, the conditional optimal transport costs will assign most unaligned samples to majority prototypes, so the 6.97% margin should shrink. A direct test is to run the same protocol on label-confusable classes such as Fashion-MNIST pullover versus coat.
  • Editorial inference: the convergence proof depends on the regularized local loss being $\alpha$-strongly convex with $\alpha = \varphi - L > 0$, but the experiments set $\varphi = 0.1$ without measuring the Lipschitz constant of the deep extractor; checking the smallest Hessian eigenvalue along training would tell whether the guarantee actually holds for the deployed configuration.
  • Editorial inference: the privacy analysis leaves residual label leakage through prototype similarity; the paper's own noise experiments show accuracy erodes at $\kappa = 0.05$, so there is a direct trade-off curve that could be turned into a practical guideline for choosing the noise scale given a target accuracy.
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 / 8 minor

Summary. The paper proposes Proto-EVFL, a vertical federated learning (VFL) framework designed for settings where aligned labeled samples are scarce and each party holds large amounts of unaligned, unlabeled, and class-imbalanced local data. The method introduces per-party class prototypes, a probabilistic dual transport cost (PDTC) for selecting unaligned samples, a mixed-prior module that combines local and global class priors, and an adaptive gated feature aggregation strategy for the active party's classifier. The authors claim that this is the first bi-level optimization framework in VFL and provide a convergence analysis with a 1/sqrt(T) rate, plus extensive experiments on image and tabular datasets showing state-of-the-art accuracy, including in few-shot and zero-shot class-imbalanced scenarios.

Significance. If the claims are correct, the paper would make a substantial contribution: it addresses a practical and under-studied problem (extremely unaligned, class-imbalanced data in VFL) and reports consistent empirical gains in ablation and robustness experiments (Tables VI and VIII) over strong baselines such as FedHSSL. The proposed adaptive gating and mixed-prior components are intuitive and the ablations suggest each contributes. The claimed convergence guarantee for a bi-level VFL formulation is also of theoretical interest. However, the paper's central zero-shot claim is not supported by the algorithm as written, and the convergence-rate statement contains a logical gap. These issues must be resolved before the contribution can be assessed fairly. The work does not ship code, but the reported experiments are detailed enough to be reproduced.

major comments (4)
  1. [§IV-D, Eq. (15); §V-E, Table VIII] The zero-shot mechanism is not defined. In Eq. (15), the prototype update divides by N_{a,z}, the number of aligned samples of class z; for a zero-shot (unseen) class, N_{a,z}=0, so the update is undefined. In addition, the active classifier is trained with cross-entropy (Eq. (14)) on Y_a, which contains no examples of the unseen class, so a standard softmax head has no trained output weight for that class. The paper does not specify the test-time classification rule (e.g., nearest-prototype matching instead of the trained classifier), nor does it explain how prototypes for unseen classes are initialized or updated. Therefore the abstract's claim of predicting unseen classes and the reported improvements in Tables IV and VIII are not derivable from the described algorithm. The protocol for zero-shot evaluation must be specified precisely, including prototype initialization and the decision rule at test time, or the zero-shot claims should be removed.
  2. [§IV-E, text after Theorem 1] The statement 'If we choose τ = 1 in Theorem 1, we can see our convergence rate is O(1/√T)' is not supported by Eq. (17). With τ = 1, the first term is proportional to 1/T and the remainder is O(ε), giving O(1/T + ε), not O(1/√T). The O(1/√T) rate would require choosing ε = O(1/√T) (and consequently B = O(T^{1/4}), τ = O(log T)), which is not stated in the theorem. Please correct the rate derivation or state explicitly the relation between ε, T, and the bound.
  3. [§IV-E, Lemma 1 and Theorem 1; §V-A, Table I] The strong convexity condition in Lemma 1 requires α = φ − L > 0, where L is the Lipschitz constant of the local loss gradient with respect to the extractor parameters. In the experiments, φ is set to 0.1 (Table I). For the LeNet-5 and MLP extractors used, the Lipschitz constant L is unlikely to be below 0.1, and the paper provides no verification of this condition. Thus, the convergence guarantee in Theorem 1 may not apply to the reported experimental configurations. Please verify the condition for the used architectures, or revise the theory (or the regularizer strength) so that the assumptions are compatible with the experiments.
  4. [§IV-B, beginning of the section; §V-E] The method relies on the cluster assumption: 'the party's data distribution consists of separated clusters, with samples in a cluster likely sharing the same labels.' This assumption is load-bearing for the PDTC pseudo-labeling scheme (Eqs. 5–7) and for the reported gains in rare and unseen classes. The paper provides no empirical check of cluster purity in the unaligned local data, nor any analysis of how performance degrades when class clusters overlap. Please report a diagnostic (e.g., cluster purity on unaligned features) or include experiments on a dataset/configurations that violate the separated-cluster assumption, so that the scope and limitations of the method are clear.
minor comments (8)
  1. [§IV-E, Eq. (39) and Appendix C, Eq. (75)] The displayed inequality appears to be misprinted: it should be an additive bound of the form −(η′/2 − η′²L0)∥∇F∥² + (η′/2 + η′²L0)∥g − ∇F∥², not a product of the two norms. As written, the inequality is dimensionally inconsistent, although the subsequent bound in Eq. (40) uses the correct additive structure.
  2. [§IV-B and §IV-C, Eqs. (4), (6), (9)–(11)] The notation for class priors is inconsistent: p(µ_m_z), P(µ_m,t_z), and P(µ^{m,t+1}_z) are used interchangeably, and Eq. (11) has the same symbol P(µ^{m,t+1}_z) on both sides, which is confusing. Please unify the notation and clarify the update index.
  3. [§IV-C] The definition of γ_m is unclear: 'we define γ_m as N_m_zhat / N_m according to the number of local minority class z-hat.' No equation is given, and the text does not explain how γ_m is computed adaptively at each round from the PDTC classification results. Please provide a precise formula.
  4. [Figures 5–7] The figures in the manuscript are rendered as garbled text (e.g., '/uni00000015/uni00000013...'), making the communication-cost and convergence plots unreadable. Please regenerate these figures.
  5. [§V-E, Table VIII] The caption says few-shot and zero-shot experiments were conducted on categories 1 through 5, but does not define the number of samples used in the few-shot setting or the exact protocol for zero-shot (i.e., whether the class is absent from all aligned data or only from the active party's aligned data). Please clarify.
  6. [Algorithm 1, line 19] Line 19 updates 'prototypes for each party via Eq.15', but this step is inside the active party's loop and uses only aligned representations. The unaligned data from passive parties are never used in this update. Please clarify whether prototypes are updated only with aligned data, and how this interacts with the zero-shot setting.
  7. [§IV-B, Eq. (5)] The expectation notation 'E f m,n u ∼ f mu' is malformed; it should be written as E_{f^{m,n}_u ∼ f^m_u} [...], and the distribution f^m_u should be defined explicitly.
  8. [Throughout] The acronyms 'PDTC' and 'PTDC' are used interchangeably (e.g., 'PTDC' appears where 'PDTC' is meant); please standardize the terminology.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the convergence rate is imported from external bilevel lemmas and the prototype/prior loop is evaluated on held-out labels.

full rationale

I examined the claimed derivation chain. The bi-level objective (Eqs. 2-3) and the local/global losses (Eqs. 4-8, 14) are stated independently of the convergence target. Theorem 1's 1/sqrt(T) rate follows from Assumptions 1-5 plus Lemma 3 (Lemma 2.2 of [57]) and Lemma 4 (Lemma 9 of [60]); both lemmas are attributed to external works and are not results of this paper, so importing them is not self-citation. The prototype update (Eq. 15) and prior/prototype selection loop (Eqs. 9-11) constitute a self-training procedure, but the paper's accuracy claims are measured against held-out labels rather than read off from the fitted prototypes or priors, so no reported 'prediction' is equal to an input by construction. No parameter is fitted to the Table IV/VIII numbers and then renamed a prediction. I found no load-bearing self-citation, no imported uniqueness theorem from the authors, and no ansatz smuggled in via citation. The omission of the proofs of Lemmas 1-2 ('As the proof of Lemma 1 and 2 is too trivial, we omit it') is an omitted-proof concern, not circularity. The zero-shot setting is a genuine correctness risk: Eq. 15 divides by N_{a,z}, which is zero for an unseen class, and the cross-entropy head has no trained weight for that class; however, an undefined or undocumented mechanism is a correctness gap, not a reduction of a result to its own input. Therefore the derivation chain is self-contained with respect to circularity, and the score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The method rests on four kinds of unproved inputs: hand-set hyperparameters (φ, ρ, γ_m, learning rates) whose values are not derived from any principle; the cluster assumption that unaligned samples group into label-homogeneous clusters; the assumption that parties' unaligned data share one label space (Eq 1); and the standard regularity assumptions of bilevel analysis (Assumptions 1-5), which are cited from [57]-[59] and cannot be verified for deep extractors. Lemma 1's strong convexity condition φ > L is especially fragile because experiments fix φ = 0.1. No new physical or mathematical entities are invented: class prototypes are data-derived vectors whose use is ablated, so they carry independent empirical evidence.

free parameters (4)
  • φ (local extractor L2 regularization) = 0.1
    Hand-set in Tab I. The convergence proof's Lemma 1 requires φ > L for strong convexity of the local loss, so the theory applies only if the local loss's Lipschitz constant L is below 0.1, which is unverified for the LeNet-5/MLP extractors.
  • ρ (prototype update speed) = 0.1
    Hand-set in Tab I; controls how fast class prototypes are moved toward aligned features in Eq 15, directly shaping the PDTC selection in the next round.
  • γ_m (global prior influence ratio) = N^m_ẑ / N^m (heuristic), described as adaptively computed
    Sec IV-C gives the ratio of minority-class count to total count, but also says it is adapted from PDTC classification results each round. The ambiguity matters because γ_m controls the mix in Eq 11 and hence the whole prior-guided selection.
  • Learning rates η, η' = 1e-5 (extractor), 5e-4 or 3e-3 (active side)
    Hand-set in Tab I. The convergence theorem places conditions on step sizes (η = 2/(L2+α), η' < 1/(2L0)) that do not match the reported constants, so the proof does not govern the implemented schedule.
assumptions (5)
  • domain assumption Cluster assumption: each party's data distribution consists of separated clusters whose samples likely share labels.
    Stated explicitly in Sec IV-B, it underwrites the PDTC loss (Eqs 5-7) and the EM prior estimation (Appendix B2). If unaligned samples do not form label-homogeneous clusters, pseudo-label assignment reinforces majority classes and the claimed gains collapse.
  • domain assumption Label-space consistency of unaligned data across parties (Y^m_{a+u} = Y^{m+1}_{a+u}, Eq 1).
    The mixed global prior and the shared prototype set assume all parties' unaligned samples live in the same label space, including rare and unseen classes; violated if parties hold disjoint class sets.
  • ad hoc to paper Strong convexity of the local loss in E (Lemma 1) with α = φ - L > 0.
    Required for the 1/√T theorem. The implemented local loss is a deep-network loss that is not verifiably strongly convex, and the experimental φ = 0.1 is not checked against L; the theorem therefore applies to a stylized objective, not the network actually trained.
  • standard math Assumptions 1-5: Lipschitz, smoothness, second-derivative Lipschitz, bounded variance, bounded domain.
    Standard regularity conditions imported from [57], [58], [59]. They are canonical for bilevel SGD analysis, but unverifiable for deep extractors and gating networks.
  • domain assumption Trustworthy active party for privacy claims (Sec IV-F).
    The privacy analysis assumes the active party is honest and that small noise (κ < 0.05 per Tabs XII and XIII) preserves accuracy; the paper's own experiments show stronger noise degrades accuracy below Vanilla VFL.
invented entities (1)
  • Per-party class prototypes μ^m_z independent evidence
    purpose: Anchor the dual optimal transport selection, the EM prior estimation, and the claimed unseen-class prediction capability.
    Prototypes are computed from aligned data via Eq 15 and their contribution is ablated in Tab VI, so they carry direct empirical evidence; however, their initialization is never specified, which matters because unseen-class prototypes never receive updates (N_{a,z}=0).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Proto-EVFL: Enhanced Vertical Federated Learning via Dual Prototype with Extremely Unaligned Data." pith.science (2026). https://pith.science/paper/2GSK6SRS

@misc{pith2026250722488,
  author       = {Pith},
  title        = {Pith review of: Proto-EVFL: Enhanced Vertical Federated Learning via Dual Prototype with Extremely Unaligned Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2GSK6SRS}},
  note         = {Machine review of arXiv:2507.22488}
}
read the original abstract

In vertical federated learning (VFL), multiple enterprises address aligned sample scarcity by leveraging massive locally unaligned samples to facilitate collaborative learning. However, unaligned samples across different parties in VFL can be extremely class-imbalanced, leading to insufficient feature representation and limited model prediction space. Specifically, class-imbalanced problems consist of intra-party class imbalance and inter-party class imbalance, which can further cause local model bias and feature contribution inconsistency issues, respectively. To address the above challenges, we propose Proto-EVFL, an enhanced VFL framework via dual prototypes. We first introduce class prototypes for each party to learn relationships between classes in the latent space, allowing the active party to predict unseen classes. We further design a probabilistic dual prototype learning scheme to dynamically select unaligned samples by conditional optimal transport cost with class prior probability. Moreover, a mixed prior guided module guides this selection process by combining local and global class prior probabilities. Finally, we adopt an \textit{adaptive gated feature aggregation strategy} to mitigate feature contribution inconsistency by dynamically weighting and aggregating local features across different parties. We proved that Proto-EVFL, as the first bi-level optimization framework in VFL, has a convergence rate of 1/\sqrt T. Extensive experiments on various datasets validate the superiority of our Proto-EVFL. Even in a zero-shot scenario with one unseen class, it outperforms baselines by at least 6.97%

Figures

Figures reproduced from arXiv: 2507.22488 by the authors.

Figure 1
Figure 1. Intra-party class imbalance and inter-party class imbalance in VFL. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Model bias and inconsistent feature contribution issues in the VFL. The purple, white, green, and red represent the different classes, respectively. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The overview framework of Proto-EVFL. The diagram shows the active party on a blue background and the passive party on a yellow background. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Comparison of different methods for classifying unaligned and [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Communication cost comparisons on four datasets. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Convergence comparisons on four datasets. [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Communication round comparisons on four datasets. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

88 extracted references · 69 canonical work pages

  1. [57]

    Approximation methods for bilevel program- ming,

    S. Ghadimi and M. Wang, “Approximation methods for bilevel program- ming,” arXiv preprint arXiv:1802.02246 , 2018

    arXiv 2018

  2. [60]

    Bilevel optimization: Convergence analysis and enhanced design,

    K. Ji, J. Yang, and Y . Liang, “Bilevel optimization: Convergence analysis and enhanced design,” in International conference on machine learning. PMLR, 2021, pp. 4882–4892

    work page 2021

  3. [1]

    Communication-efficient learning of deep networks from decentralized data,

    H. B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” in International Conference on Artificial Intelligence and Statistics, 2016

    2016

  4. [2]

    Federated learning for privacy- preserving ai,

    Y . Cheng, Y . Liu, T. Chen, and Q. Yang, “Federated learning for privacy- preserving ai,” Communications of the ACM, vol. 63, no. 12, pp. 33–36, 2020

    2020

  5. [3]

    Federated machine learning: Concept and applications,

    Q. Yang, Y . Liu, T. Chen, and Y . Tong, “Federated machine learning: Concept and applications,” ACM Transactions on Intelligent Systems and Technology (TIST), vol. 10, no. 2, pp. 1–19, 2019

    2019

  6. [4]

    Vertical federated learning,

    Y . Liu, Y . Kang, T. Zou, Y . Pu, Y . He, X. Ye, Y . Ouyang, Y .- Q. Zhang, and Q. Yang, “Vertical federated learning,” arXiv preprint arXiv:2211.12814, 2022

    arXiv 2022

  7. [5]

    Federated learning in mobile edge networks: A comprehensive survey,

    W. Y . B. Lim, N. C. Luong, D. T. Hoang, Y . Jiao, Y .-C. Liang, Q. Yang, D. Niyato, and C. Miao, “Federated learning in mobile edge networks: A comprehensive survey,” IEEE Communications Surveys & Tutorials , vol. 22, no. 3, pp. 2031–2063, 2020

    2020

  8. [6]

    Communication-efficient vertical federated learning with limited overlapping samples,

    J. Sun, Z. Xu, D. Yang, V . Nath, W. Li, C. Zhao, D. Xu, Y . Chen, and H. R. Roth, “Communication-efficient vertical federated learning with limited overlapping samples,” arXiv preprint arXiv:2303.16270 , 2023

    arXiv 2023

Show all 88 references
  1. [7]

    Fedcvt: Semi-supervised vertical federated learning with cross-view training,

    Y . Kang, Y . Liu, and X. Liang, “Fedcvt: Semi-supervised vertical federated learning with cross-view training,” ACM Transactions on Intelligent Systems and Technology (TIST) , vol. 13, no. 4, pp. 1–16, 2022

    2022

  2. [8]

    Vertical federated learning: Challenges, methodologies and experiments,

    K. Wei, J. Li, C. Ma, M. Ding, S. Wei, F. Wu, G. Chen, and T. Ranbaduge, “Vertical federated learning: Challenges, methodologies and experiments,” arXiv preprint arXiv:2202.04309 , 2022

    2022 arXiv

  3. [9]

    Semi-supervised federated heterogeneous transfer learning,

    S. Feng, B. Li, H. Yu, Y . Liu, and Q. Yang, “Semi-supervised federated heterogeneous transfer learning,” Knowledge-Based Systems , vol. 252, p. 109384, 2022

    2022

  4. [10]

    A hybrid self-supervised learning framework for vertical federated learning,

    Y . He, Y . Kang, X. Zhao, J. Luo, L. Fan, Y . Han, and Q. Yang, “A hybrid self-supervised learning framework for vertical federated learning,” arXiv preprint arXiv:2208.08934 , 2022

    2022 arXiv

  5. [11]

    Self- supervised cross-silo federated neural architecture search,

    X. Liang, Y . Liu, J. Luo, Y . He, T. Chen, and Q. Yang, “Self- supervised cross-silo federated neural architecture search,” arXiv preprint arXiv:2101.11896, 2021

    2021 arXiv

  6. [12]

    Self-supervised vertical feder- ated learning,

    T. Castiglia, S. Wang, and S. Patterson, “Self-supervised vertical feder- ated learning,” in Workshop on Federated Learning: Recent Advances and New Challenges (in Conjunction with NeurIPS 2022) , 2022

    2022

  7. [13]

    Vertical semi- federated learning for efficient online advertising,

    W. Li, Q. Xia, H. Cheng, K. Xue, and S.-T. Xia, “Vertical semi- federated learning for efficient online advertising,” arXiv preprint arXiv:2209.15635, 2022

    2022

  8. [14]

    Multi-view federated learning with data collaboration,

    Y . Yang, X. Ye, and T. Sakurai, “Multi-view federated learning with data collaboration,” in 2022 14th International Conference on Machine Learning and Computing (ICMLC) , 2022, pp. 178–183

    2022

  9. [15]

    Vertical federated learning-based feature selection with non- overlapping sample utilization,

    S. Feng, “Vertical federated learning-based feature selection with non- overlapping sample utilization,” Expert Systems with Applications , vol. 208, p. 118097, 2022

    2022

  10. [16]

    Practical vertical federated learning with unsupervised representation learning,

    Z. Wu, Q. Li, and B. He, “Practical vertical federated learning with unsupervised representation learning,” IEEE Transactions on Big Data , 2022

    2022

  11. [17]

    A review of the oversampling techniques in class imbalance problem,

    S. Sharma, A. Gosain, and S. Jain, “A review of the oversampling techniques in class imbalance problem,” in International Conference on Innovative Computing and Communications: Proceedings of ICICC 2021, Volume 1. Springer, 2022, pp. 459–472

    2021

  12. [18]

    A review on imbalanced data handling using undersampling and oversampling technique,

    M. S. Shelke, P. R. Deshmukh, and V . K. Shandilya, “A review on imbalanced data handling using undersampling and oversampling technique,” Int. J. Recent Trends Eng. Res , vol. 3, no. 4, pp. 444–449, 2017

    2017

  13. [19]

    Data selection for federated learning with relevant and irrelevant data at clients,

    T. Tuor, S. Wang, B. Ko, C. Liu, and K. K. Leung, “Data selection for federated learning with relevant and irrelevant data at clients,” ArXiv, vol. abs/2001.08300, 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:210861211

    2001 arXiv

  14. [20]

    Three approaches for personalization with applications to federated learning,

    Y . Mansour, M. Mohri, J. Ro, and A. T. Suresh, “Three approaches for personalization with applications to federated learning,” arXiv preprint arXiv:2002.10619, 2020

    2002 arXiv

  15. [21]

    Attribute-based classifi- cation for zero-shot visual object categorization,

    C. H. Lampert, H. Nickisch, and S. Harmeling, “Attribute-based classifi- cation for zero-shot visual object categorization,” IEEE transactions on pattern analysis and machine intelligence , vol. 36, no. 3, pp. 453–465, 2013

    2013

  16. [22]

    Zero-data learning of new tasks

    H. Larochelle, D. Erhan, and Y . Bengio, “Zero-data learning of new tasks.” in AAAI, vol. 1, no. 2, 2008, p. 3

    2008

  17. [23]

    Learning hypergraph-regularized attribute predictors,

    S. Huang, M. Elhoseiny, A. Elgammal, and D. Yang, “Learning hypergraph-regularized attribute predictors,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2015, pp. 409– 417

    2015

  18. [24]

    Learning multimodal latent attributes,

    Y . Fu, T. M. Hospedales, T. Xiang, and S. Gong, “Learning multimodal latent attributes,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 36, no. 2, pp. 303–316, 2014

    2014

  19. [25]

    Zero-shot recognition with unreliable attributes,

    D. Jayaraman and K. Grauman, “Zero-shot recognition with unreliable attributes,” in Advances in Neural Information Processing Systems, Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, and K. Weinberger, Eds., vol. 27. Curran Associates, Inc., JOURNAL OF LATEX CLASS FILES, VO...

    2024

  20. [26]

    Attribute-based classifi- cation for zero-shot visual object categorization,

    C. H. Lampert, H. Nickisch, and S. Harmeling, “Attribute-based classifi- cation for zero-shot visual object categorization,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 36, no. 3, pp. 453–465, 2014

    2014

  21. [27]

    Generative zero-shot learning via low- rank embedded semantic dictionary,

    Z. Ding, M. Shao, and Y . Fu, “Generative zero-shot learning via low- rank embedded semantic dictionary,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 41, no. 12, pp. 2861–2874, 2019

    2019

  22. [28]

    Zero-shot learning via latent space encoding,

    Y . Yu, Z. Ji, J. Guo, and Z. Zhang, “Zero-shot learning via latent space encoding,” IEEE Transactions on Cybernetics, vol. 49, no. 10, pp. 3755– 3766, 2019

    2019

  23. [29]

    Transduc- tive zero-shot learning with a self-training dictionary approach,

    Y . Yu, Z. Ji, X. Li, J. Guo, Z. Zhang, H. Ling, and F. Wu, “Transduc- tive zero-shot learning with a self-training dictionary approach,” IEEE Transactions on Cybernetics , vol. 48, no. 10, pp. 2908–2919, 2018

    2018

  24. [30]

    Feature generating networks for zero-shot learning,

    Y . Xian, T. Lorenz, B. Schiele, and Z. Akata, “Feature generating networks for zero-shot learning,” CoRR, vol. abs/1712.00981, 2017. [Online]. Available: http://arxiv.org/abs/1712.00981

    2017 arXiv

  25. [31]

    General- ized zero- and few-shot learning via aligned variational autoencoders,

    E. Schonfeld, S. Ebrahimi, S. Sinha, T. Darrell, and Z. Akata, “General- ized zero- and few-shot learning via aligned variational autoencoders,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019

    2019

  26. [32]

    Fedproto: Federated prototype learning across heterogeneous clients,

    Y . Tan, G. Long, L. Liu, T. Zhou, Q. Lu, J. Jiang, and C. Zhang, “Fedproto: Federated prototype learning across heterogeneous clients,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 8, 2022, pp. 8432–8440

    2022

  27. [33]

    Fedfa: Federated learning with feature anchors to align features and classifiers for heterogeneous data,

    T. Zhou, J. Zhang, and D. H. Tsang, “Fedfa: Federated learning with feature anchors to align features and classifiers for heterogeneous data,” IEEE Transactions on Mobile Computing , 2023

    2023

  28. [34]

    Personalized federated learning with feature alignment and classifier collaboration,

    J. Xu, X. Tong, and S.-L. Huang, “Personalized federated learning with feature alignment and classifier collaboration,” arXiv preprint arXiv:2306.11867, 2023

    2023 arXiv

  29. [35]

    Tackling data heterogeneity in federated learning with class prototypes,

    Y . Dai, Z. Chen, J. Li, S. Heinecke, L. Sun, and R. Xu, “Tackling data heterogeneity in federated learning with class prototypes,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 37, no. 6, 2023, pp. 7314–7322

    2023

  30. [36]

    Fedproc: Prototypical contrastive federated learning on non-iid data,

    X. Mu, Y . Shen, K. Cheng, X. Geng, J. Fu, T. Zhang, and Z. Zhang, “Fedproc: Prototypical contrastive federated learning on non-iid data,” Future Generation Computer Systems , vol. 143, pp. 93–104, 2023

    2023

  31. [37]

    Contrastive-enhanced domain generalization with federated learning,

    X. Yu, D. Wang, M. McKeown, and Z. J. Wang, “Contrastive-enhanced domain generalization with federated learning,” IEEE Transactions on Artificial Intelligence, 2023

    2023

  32. [38]

    Vertical federated knowledge trans- fer via representation distillation for healthcare collaboration networks,

    C.-j. Huang, L. Wang, and X. Han, “Vertical federated knowledge trans- fer via representation distillation for healthcare collaboration networks,” in Proceedings of the ACM Web Conference 2023, 2023, pp. 4188–4199

    2023

  33. [39]

    Improving availability of vertical federated learning: Relaxing inference on non-overlapping data,

    Z. Ren, L. Yang, and K. Chen, “Improving availability of vertical federated learning: Relaxing inference on non-overlapping data,” ACM Transactions on Intelligent Systems and Technology (TIST) , vol. 13, no. 4, pp. 1–20, 2022

    2022

  34. [40]

    Entity resolution and federated learning get a federated resolution,

    R. Nock, S. Hardy, W. Henecka, H. Ivey-Law, G. Patrini, G. Smith, and B. Thorne, “Entity resolution and federated learning get a federated resolution,” arXiv preprint arXiv:1803.04035 , 2018

    2018 arXiv

  35. [41]

    Private federated learning on vertically partitioned data via entity resolution and additively homomorphic encryption,

    S. Hardy, W. Henecka, H. Ivey-Law, R. Nock, G. Patrini, G. Smith, and B. Thorne, “Private federated learning on vertically partitioned data via entity resolution and additively homomorphic encryption,” arXiv preprint arXiv:1711.10677, 2017

    2017 arXiv

  36. [42]

    Opti- mal transport based one-shot federated learning for artificial intelligence of things,

    Y .-H. Chiang, K. Terai, T.-W. Chiang, H. Lin, Y . Ji, and J. C. Lui, “Opti- mal transport based one-shot federated learning for artificial intelligence of things,” IEEE Internet of Things Journal , 2023

    2023

  37. [43]

    Global and local prompts coop- eration via optimal transport for federated learning,

    H. Li, W. Huang, J. Wang, and Y . Shi, “Global and local prompts coop- eration via optimal transport for federated learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 12 151–12 161

    2024

  38. [44]

    Spectr: Fast speculative decoding via optimal transport,

    Z. Sun, A. T. Suresh, J. H. Ro, A. Beirami, H. Jain, and F. Yu, “Spectr: Fast speculative decoding via optimal transport,” Advances in Neural Information Processing Systems , vol. 36, 2024

    2024

  39. [45]

    Bayes’ theorem,

    J. Joyce, “Bayes’ theorem,” 2003

    2003

  40. [46]

    Minimal-entropy correlation alignment for unsupervised deep domain adaptation,

    P. Morerio, J. Cavazza, and V . Murino, “Minimal-entropy correlation alignment for unsupervised deep domain adaptation,” arXiv preprint arXiv:1711.10288, 2017

    2017 arXiv

  41. [47]

    Entropy minimization vs. diversity maximization for domain adaptation,

    X. Wu, Q. Zhou, Z. Yang, C. Zhao, L. J. Latecki et al. , “Entropy minimization vs. diversity maximization for domain adaptation,” arXiv preprint arXiv:2002.01690, 2020

    2002 arXiv

  42. [48]

    von liebig’s law of the minimum and plankton ecology (1899–1991),

    H. De Baar, “von liebig’s law of the minimum and plankton ecology (1899–1991),” Progress in oceanography , vol. 33, no. 4, pp. 347–386, 1994

    1991

  43. [49]

    Enhancing supervised learning with unlabeled data,

    S. Goldman and Y . Zhou, “Enhancing supervised learning with unlabeled data,” in ICML. Citeseer, 2000, pp. 327–334

    2000

  44. [50]

    Adjusting the outputs of a classifier to new a priori probabilities: a simple procedure,

    M. Saerens, P. Latinne, and C. Decaestecker, “Adjusting the outputs of a classifier to new a priori probabilities: a simple procedure,” Neural computation, vol. 14, no. 1, pp. 21–41, 2002

    2002

  45. [51]

    A unified solution for privacy and communication efficiency in vertical federated learning,

    G. Wang, B. Gu, Q. Zhang, X. Li, B. Wang, and C. X. Ling, “A unified solution for privacy and communication efficiency in vertical federated learning,” Advances in Neural Information Processing Systems , vol. 36, 2024

    2024

  46. [52]

    Flexible vertical federated learning with heterogeneous parties,

    T. Castiglia, S. Wang, and S. Patterson, “Flexible vertical federated learning with heterogeneous parties,” IEEE Transactions on Neural Networks and Learning Systems , 2023

    2023

  47. [53]

    Twenty years of mixture of experts,

    S. E. Yuksel, J. N. Wilson, and P. D. Gader, “Twenty years of mixture of experts,” IEEE transactions on neural networks and learning systems , vol. 23, no. 8, pp. 1177–1193, 2012

    2012

  48. [54]

    Less-vfl: Communication-efficient feature selection for vertical federated learning,

    T. Castiglia, Y . Zhou, S. Wang, S. Kadhe, N. Baracaldo, and S. Pat- terson, “Less-vfl: Communication-efficient feature selection for vertical federated learning,” in International Conference on Machine Learning . PMLR, 2023, pp. 3757–3781

    2023

  49. [55]

    Label inference attacks against vertical federated learning,

    C. Fu, X. Zhang, S. Ji, J. Chen, J. Wu, S. Guo, J. Zhou, A. X. Liu, and T. Wang, “Label inference attacks against vertical federated learning,” in 31st USENIX security symposium (USENIX Security 22) , 2022, pp. 1397–1414

    2022

  50. [56]

    Practical feature inference attack in vertical federated learning during prediction in artificial internet of things,

    R. Yang, J. Ma, J. Zhang, S. Kumari, S. Kumar, and J. J. Rodrigues, “Practical feature inference attack in vertical federated learning during prediction in artificial internet of things,” IEEE Internet of Things Journal, vol. 11, no. 1, pp. 5–16, 2023

    2023

  51. [58]

    Convergence of meta- learning with task-specific adaptation over partial parameters,

    K. Ji, J. D. Lee, Y . Liang, and H. V . Poor, “Convergence of meta- learning with task-specific adaptation over partial parameters,” Advances in Neural Information Processing Systems , vol. 33, pp. 11 490–11 500, 2020

    2020

  52. [59]

    Closing the convergence gap of sgd without replacement,

    S. Rajput, A. Gupta, and D. Papailiopoulos, “Closing the convergence gap of sgd without replacement,” in International Conference on Ma- chine Learning. PMLR, 2020, pp. 7964–7973

    2020

  53. [61]

    Fastslowmo: Federated learning with combined worker and aggregator momenta,

    Z. Yang, S. Fu, W. Bao, D. Yuan, and A. Y . Zomaya, “Fastslowmo: Federated learning with combined worker and aggregator momenta,” IEEE Transactions on Artificial Intelligence , vol. 4, no. 5, pp. 1041– 1050, 2022

    2022

  54. [62]

    General data protection regulation,

    G. D. P. R. GDPR, “General data protection regulation,” URL: https://gdpr-info. eu/[accessed 2020-11-21] , 2018

    2020

  55. [63]

    Inverting gradients-how easy is it to break privacy in federated learning?

    J. Geiping, H. Bauermeister, H. Dr ¨oge, and M. Moeller, “Inverting gradients-how easy is it to break privacy in federated learning?” Ad- vances in neural information processing systems , vol. 33, pp. 16 937– 16 947, 2020

    2020

  56. [64]

    3d shapenets: A deep representation for volumetric shapes,

    Z. Wu, S. Song, A. Khosla, F. Yu, L. Zhang, X. Tang, and J. Xiao, “3d shapenets: A deep representation for volumetric shapes,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2015, pp. 1912–1920

    2015

  57. [65]

    Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,

    H. Xiao, K. Rasul, and R. V ollgraf, “Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,” 08 2017

    2017

  58. [66]

    Default of Credit Card Clients,

    I.-C. Yeh, “Default of Credit Card Clients,” UCI Machine Learning Repository, 2016, DOI: https://doi.org/10.24432/C55S3H

    2016 doi

  59. [67]

    Becker and R

    B. Becker and R. Kohavi, “Adult,” UCI Machine Learning Repository, 1996, DOI: https://doi.org/10.24432/C5XW20

    1996 doi

  60. [68]

    A method for stochastic optimization,

    D. Kinga, J. B. Adam et al. , “A method for stochastic optimization,” in International conference on learning representations (ICLR) , vol. 5. San Diego, California;, 2015, p. 6

    2015

  61. [69]

    An experimental study of class imbalance in federated learning,

    C. Xiao and S. Wang, “An experimental study of class imbalance in federated learning,” in 2021 IEEE Symposium Series on Computational Intelligence (SSCI). IEEE, 2021, pp. 1–7

    2021

  62. [70]

    Semantic cosine similar- ity,

    F. Rahutomo, T. Kitasuka, M. Aritsugi et al., “Semantic cosine similar- ity,” in The 7th international student conference on advanced science and technology ICAST, vol. 4, no. 1. University of Seoul South Korea, 2012, p. 1

    2012

  63. [71]

    Learning with a wasserstein loss,

    C. Frogner, C. Zhang, H. Mobahi, M. Araya, and T. A. Poggio, “Learning with a wasserstein loss,” Advances in neural information processing systems, vol. 28, 2015

    2015

  64. [72]

    Semi-supervised cross-silo advertising with partial knowledge transfer,

    W. Li, Q. Xia, J. Deng, H. Cheng, J. Liu, K. Xue, Y . Cheng, and S.- T. Xia, “Semi-supervised cross-silo advertising with partial knowledge transfer,” arXiv e-prints, pp. arXiv–2205, 2022

    2022

  65. [73]

    Differential privacy,

    C. Dwork, “Differential privacy,” in International colloquium on au- tomata, languages, and programming . Springer, 2006, pp. 1–12. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2024 15

    2006

  66. [74]

    Semi-supervised learning by entropy minimization,

    Y . Grandvalet and Y . Bengio, “Semi-supervised learning by entropy minimization,” Advances in neural information processing systems , vol. 17, 2004

    2004

  67. [75]

    Semi- supervised domain adaptation via minimax entropy,

    K. Saito, D. Kim, S. Sclaroff, T. Darrell, and K. Saenko, “Semi- supervised domain adaptation via minimax entropy,” in Proceedings of the IEEE/CVF international conference on computer vision , 2019, pp. 8050–8058

    2019

  68. [76]

    Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation,

    T.-H. Vu, H. Jain, M. Bucher, M. Cord, and P. P ´erez, “Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2019, pp. 2517–2526

    2019

  69. [77]

    Universal domain adaptation through self supervision,

    K. Saito, D. Kim, S. Sclaroff, and K. Saenko, “Universal domain adaptation through self supervision,” Advances in neural information processing systems, vol. 33, pp. 16 282–16 292, 2020

    2020

  70. [78]

    Robust optimal transport with applications in generative modeling and domain adaptation,

    Y . Balaji, R. Chellappa, and S. Feizi, “Robust optimal transport with applications in generative modeling and domain adaptation,” Advances in Neural Information Processing Systems , vol. 33, pp. 12 934–12 944, 2020

    2020

  71. [79]

    G. J. McLachlan, Discriminant analysis and statistical pattern recogni- tion. John Wiley & Sons, 2005

    2005

  72. [80]

    Cross-silo federated neural architecture search for heterogeneous and cooperative systems,

    Y . Liu, X. Liang, J. Luo, Y . He, T. Chen, Q. Yao, and Q. Yang, “Cross-silo federated neural architecture search for heterogeneous and cooperative systems,” in Federated and Transfer Learning . Springer, 2022, pp. 57–86. Wei Guo She is a Ph.D. student at the Insti- tute of Ar...

    2022

  73. [82]

    Likelihood

    Connecting with Entropy Minimization: To understand more clearly the expected cost from feature representation to class prototypes, we can think of it as an extension of entropy minimization [74], an effective form of regularization that is widely used in many previous prior d...

    2024

  74. [83]

    The target label of each data point, Y m,n, can not be directly obtained by observation

    Local prior probability estimation: To estimate local class proportions {P (µm,t z )}Z z=1 of party m in round t, we adopt the EM algorithm to infer the iterative updates of them. The target label of each data point, Y m,n, can not be directly obtained by observation. We view ...

  75. [84]

    To show the smooth property of F (Θ), we first introduce the following lemma which is proposed in [57]

    (57) As the proof of Lemma 1 and 2 is too trivial, we omit it. To show the smooth property of F (Θ), we first introduce the following lemma which is proposed in [57]. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2024 19 Lemma 3. (Lemma 2.2 in [57]) Under Assumptions 1,...

    2024

  76. [85]

    Other nota- tions involved Bj or Bj such as ∇Θ∇E llocal(Θ0 t , E j−1 t ; Bj−1) have similar meanings

    For the sake of clarity, we introduce a new notation ∂lglobal(Θ0 t ,E τ t ;B0) ∂Θ0 t := 1 B P i∈B0 ∂lglobal(Θ0 t ,E τ t ;ξi) ∂Θ0 t . Other nota- tions involved Bj or Bj such as ∇Θ∇E llocal(Θ0 t , E j−1 t ; Bj−1) have similar meanings. Proposition 1. Under Assumptions 1-5, choo...

    2024

  77. [86]

    (77) This is a more general result of Theorem 1

    − inf Θ F (Θ) τ T( η′ 2 − L0η′2) + (η′ + 2η′2L0)L2 2∆2 τ − 1 τ + (η′ + 2η′2L0)τ {(L2 + L2 2 α ) h L2 − α L2 + α τ √ ∆ + σ√L2αB ] + L1[ L2(1 − 2 L2+α α)τ α + 1 α √ B ( L2 2 α + L2) + σ α√L2αB ( L2L4 α + L3) + 2 L2 + α ( L2L4 α + L3) √ ∆ (1 − 2 L2+α α)τ 1 − 2 L2+α α − L2−α L2+α ...

    2024

  78. [87]

    Datasets: To fairly evaluate the effectiveness of our proposed Proto-EVFL, we conduct experiments on two widely used image datasets, namely ModelNet-10 [64] and Fashion- MNIST [65], and two popular tabular datasets in VFL, in- cluding Credit [66] and Adult [67], as shown in Ta...

  79. [88]

    nc represents the number of samples in class z

    Class imbalanced metric: The calculation formula of MID is as follows: MID = LRID LRIDextreme = logZ N ZX z=1 nz ln N Zn z , (78) where N refers to the total number of datasets. nc represents the number of samples in class z. LRID can be obtained by: LRID = −2 ZX z=1 nz ln N Z...

  80. [2014]

    Available: https://proceedings.neurips.cc/paper files/ paper/2014/file/1f1baa5b8edac74eb4eaa329f14a0361-Paper.pdf

    [Online]. Available: https://proceedings.neurips.cc/paper files/ paper/2014/file/1f1baa5b8edac74eb4eaa329f14a0361-Paper.pdf

    2014

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.