Pith. sign in

REVIEW 2 major objections 5 minor 64 references

Where synthetic samples go in federated learning matters more than how many are generated; an entropy-adaptive budget recovers most accuracy at 94% less cost.

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 ↗

T0 review · grok-4.5

2026-07-11 01:37 UTC pith:24E54YLF

load-bearing objection Clean allocation paper: entropy-adaptive fill-to-threshold recovers most of Full-Balance accuracy at ~6% of the samples, and matched-budget gaps show WHERE really matters. the 2 major comments →

arxiv 2607.06616 v1 pith:24E54YLF submitted 2026-07-07 cs.LG cs.AIcs.CV

WHERE to Generate Matters: Budget-Aware Synthetic Augmentation for Label Skewed Federated Learning

Sangwoo Lee , Sunghwan Park , Jaewoo Lee This is my paper
classification cs.LG cs.AIcs.CV
keywords Federated LearningLabel SkewSynthetic Data AugmentationBudget AllocationDiffusion ModelsClient DriftEntropy-Adaptive Budget
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.

Label skew in federated learning leaves clients with uneven class counts, so their updates drift and the global model suffers. The usual fix—generating synthetic samples until every local class is fully balanced—works but is extremely expensive. This paper shows that a simple rule based on each client’s label entropy can decide both how much that client should generate and which scarce classes the samples should fill. The resulting policy recovers most of the accuracy gain of full balancing while using 94.1% fewer synthetic samples, and at any matched total budget it beats uniform placement by a large margin. Readers care because generation cost is the practical bottleneck; allocating that budget by local entropy turns the same samples into substantially more accuracy.

Core claim

Synthetic augmentation for label-skewed federated learning is primarily a budget-allocation problem: how much each client generates and where the samples go matter more than generating until every class is balanced. FedEAS computes a per-class generation budget from each client’s normalized entropy and mean class count; skewed clients receive larger budgets, samples fill only classes below that threshold, and the total generation budget is the sum of these local decisions rather than a fixed global quota. On CIFAR-10 and CIFAR-100 under Dirichlet label skew this recovers most of the accuracy gain of full class balancing while cutting generation by 94.1%, and under matched budgets it outperfo

What carries the argument

The entropy-adaptive budget rule bk = floor(sqrt(mean class count) × (1 − normalized entropy) × β). It turns each client’s local label distribution into a per-class generation count that automatically gives more budget to skewed clients, routes samples only to scarce classes, and lets a single scalar β set the overall operating point.

Load-bearing premise

The main experiments assume a high-quality class-conditional generator trained on the entire centralized training set, which is often unavailable or privacy-prohibited in real federated deployments.

What would settle it

Replace the full-training-set generator with one trained only on private client data (or under differential privacy) and re-run the matched-budget CIFAR experiments; if FedEAS no longer recovers most of the full-balance accuracy gain at roughly 6% of the generation volume, or if its gap over uniform allocation collapses, the allocation claim does not survive realistic generators.

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

If this is right

  • Full class balancing is unnecessary; most of its accuracy gain is available at a small fraction of the generation cost.
  • At any fixed generation budget, placing samples by local label entropy substantially outperforms uniform or missing-class-only placement.
  • Accuracy rises monotonically with the budget parameter, so the operating point can be chosen from available generation resources alone.
  • In the balanced (IID) limit the policy reduces to ordinary FedAvg and can be left on by default.
  • Data-level entropy budgets can be combined with optimization-level corrections to exceed either method alone.

Where Pith is reading between the lines

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

  • The same fill-to-threshold idea may transfer to non-image federated settings whenever a class-conditional generator exists.
  • Because the rule reads only local label counts, it can sit on top of any future privacy-preserving generator without redesign.
  • The observed diminishing-returns curve of imbalance reduction suggests an automatic cost-aware stopping rule that would remove the need to hand-pick β.
  • Matched-budget protocols should become standard when comparing synthetic-augmentation policies, because total sample count alone does not determine accuracy.

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

2 major / 5 minor

Summary. The paper addresses label skew in federated learning by treating synthetic augmentation as a budget-allocation problem rather than full class balancing. FedEAS assigns each client a per-class generation budget bk = floor(sqrt(n_bar_k) * (1 - H_tilde_k) * beta) from local label counts, then fills only classes below that threshold. The total generation budget therefore emerges from the per-client rule. On CIFAR-10/100 under Dirichlet partitions the method recovers most of Full-Balance accuracy while using 94.1% fewer samples (13 437 vs 226 783 on CIFAR-10) and, under a matched-budget protocol that holds generator and total samples fixed, outperforms Uniform allocation by up to 18.82%. Supporting evidence includes multi-setting accuracy tables, a beta sweep, factor ablations, a closed-form diminishing-returns argument, and an empirical gradient-dissimilarity measurement.

Significance. If the matched-budget results hold, the work cleanly isolates allocation (WHERE) as a first-order design choice for synthetic augmentation under label skew and supplies a simple, label-count-only rule that practitioners can tune with a single scalar. The 3.7x wall-clock reduction and the transparent selection criteria for beta are concrete engineering contributions. Strengths include the matched-budget protocol that holds generator and total samples fixed, the factor ablation that isolates entropy and sub-linear size scaling, the supplementary closed-form diminishing-returns argument, and the empirical Gamma-hat measurement. The generator-realism caveat is acknowledged by the authors and partially stress-tested with SD-turbo, so the internal claim that allocation drives accuracy remains intact.

major comments (2)
  1. Section 4.1 and Limitations: the primary generator is a class-conditional DDPM trained on each dataset's full training set. While this isolates allocation from generator design and the SD-turbo transfer partially decouples the gap, absolute accuracy drops and privacy-preserving generator training remains open. A short additional experiment or clearer quantitative bound on how much of the reported gain survives under a federated or differentially-private generator would strengthen the deployment claim without altering the allocation result.
  2. Section 2.2, Eqs. (5)-(6): the link from imbalance score I to client-drift Gamma rests on three idealizing assumptions (orthogonal per-class gradients, shared gradient norm, shared per-class losses). The authors correctly note that these do not hold tightly and supply an empirical Gamma-hat measurement in the supplement; however, that measurement shows both Uniform and FedEAS nearly eliminate directional disagreement, so the 7% accuracy gap is attributed to coverage rather than alignment. A brief discussion of whether I remains a reliable surrogate once directional disagreement is already low would tighten the motivation.
minor comments (5)
  1. Figure 1(a) caption and main text: the hatched/solid bar distinction is clear once read carefully, but a short legend inside the figure would help readers who scan only the graphic.
  2. Algorithm 1 line 16: aggregation weights use original N_k; a one-sentence reminder that synthetic samples do not alter client weights would prevent a common misreading.
  3. Section 4.3: the two selection conditions for beta (remove >=50% of I(0), synthetic fraction <25%) are transparent but appear only in the text; restating them next to Figure 5(b) would improve reproducibility.
  4. Supplementary Table 2 reports peak accuracy; a single sentence in the main text noting that last-20-round means and peaks tell the same ranking story would reassure readers who prefer peak metrics.
  5. Typos: 'Accuracyalsorisesmonotonically' (page 2) and occasional missing spaces after periods in the arXiv source should be cleaned.

Circularity Check

0 steps flagged

No significant circularity: FedEAS is an explicit heuristic budget rule validated by matched-budget empirical comparisons, not a derivation that reduces to its inputs.

full rationale

The paper proposes a hand-designed allocation policy bk = floor(sqrt(n_bar_k) * (1 - H_tilde_k) * beta) motivated by three observations on an imbalance score I(B) that is computed solely from label counts (Eqs. 4-9, Section 2.2-3.1). beta is selected by two transparent, pre-training criteria on the same counts (remove >=50% of I(0) while synthetic fraction <25%; Section 4.3 and Fig. 5b). Accuracy claims are then measured on held-out global test sets under a matched-budget protocol that isolates allocation (Uniform/Missing-only/FedEAS share the identical generator and total B_FE). Nothing in the equations forces the reported test accuracies, the 94.1% budget reduction, or the gaps versus Uniform; those are empirical outcomes against independent baselines. The I-to-Gamma link is presented with explicit idealizing assumptions and is not used as a fitted predictor of accuracy. There are no self-definitional loops, no parameters fitted to the target metric and then re-reported as predictions, no load-bearing uniqueness theorems, and no self-citation chains that close the argument. The work is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

The central empirical claim rests on a small set of free design choices (β, the √n̄ scaling, the 50%/25% selection thresholds) and standard FL/domain assumptions. No new physical entities are postulated; the only invented construct is the policy itself, which is fully specified by local counts.

free parameters (3)
  • β (budget scalar) = 12
    Shared scalar that multiplies every client’s per-class budget; default β=12 chosen so that ≥50% of initial imbalance is removed while synthetic samples stay <25% of training data (Section 4.3).
  • size-factor exponent (√n̄) = 0.5
    Sub-linear scaling of budget with client size; chosen over linear n̄ to keep cross-client budget ratios moderate (Section 3.1).
  • imbalance-reduction and synthetic-fraction thresholds = 0.5 / 0.25
    50% reduction of I(0) and <25% synthetic fraction used to select the default operating point; both are hand-chosen resource criteria.
axioms (4)
  • domain assumption Label skew is the dominant source of client drift that can be mitigated by class-count rebalancing.
    Standard FL literature assumption (Section 2); used to justify intervening only on label counts.
  • domain assumption A fixed pre-trained class-conditional generator produces samples whose utility is determined primarily by allocation, not by per-sample fidelity differences across policies.
    Stated in Sections 2.1 and 6; all policies share the same generator so relative gaps are attributed to allocation.
  • ad hoc to paper First-order change in imbalance score I correctly ranks the value of placing a sample in a scarce versus abundant class.
    Derived under local linearization (Eq. 7) and used to motivate fill-to-threshold allocation (Observation 2).
  • ad hoc to paper Idealizing assumptions (orthogonal per-class gradients, shared gradient norm, shared per-class losses) link I to the FedAvg heterogeneity term Γ.
    Explicitly listed in Section 2.2; authors note they do not hold tightly and supply only an empirical ˆΓ surrogate.
invented entities (1)
  • FedEAS entropy-adaptive per-class budget rule no independent evidence
    purpose: Maps each client’s local counts to a generation budget and a fill-to-threshold allocation.
    The policy is the paper’s sole algorithmic invention; it is fully defined by Eqs. (8)–(9) and has no independent existence outside the paper.

pith-pipeline@v1.1.0-grok45 · 24206 in / 3016 out tokens · 28911 ms · 2026-07-11T01:37:10.073323+00:00 · methodology

0 comments
read the original abstract

Label skew in federated learning (FL) causes client drift and degrades global accuracy. Synthetic data augmentation can reduce this imbalance; however, full class balancing requires substantial computation cost. We propose FedEAS, a policy that assigns each client an entropy-adaptive per-class generation budget computed from its local label distribution. The budget jointly decides \emph{how much} each client generates and \emph{WHERE} the samples go. Accordingly, the total generation budget follows from the per-client budgets rather than being fixed in advance. FedEAS recovers most of the accuracy gain of full class balancing while reducing the generation budget by 94.1\%. At the same total generation budget, it outperforms Uniform allocation by up to 18.82\% across CIFAR-10 and CIFAR-100.

Figures

Figures reproduced from arXiv: 2607.06616 by Jaewoo Lee, Sangwoo Lee, Sunghwan Park.

Figure 1
Figure 1. Figure 1: Budget-aware synthetic augmentation on a CIFAR-10 Dirichlet partition (K = 20, α = 0.1). (a) Reaching full class balancing requires 226,783 synthetic samples, whereas the FedEAS(β = 12) budgets require 13,437; hatched bars show per-client full￾balancing demand, solid bars the FedEAS budgets. (b) Allocation decides how much of the budget turns into accuracy. FedAvg trains without synthetic data, while Unifo… view at source ↗
Figure 2
Figure 2. Figure 2: Imbalance concentration and reduction on the partition of [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of FedEAS. Each client summarizes its skew by the normalized en￾tropy H˜k (top left), computes the per-class generation budget bk of Eq. (8) (top right), and allocates samples only to the classes below bk (bottom). Full-Balance fills every class to its local maximum, while Uniform, Missing-only, and FedEAS place the same total generation budget differently, as compared in Section 4.2. In the IID l… view at source ↗
Figure 4
Figure 4. Figure 4: FedEAS converges faster than Missing-only under matched budgets. The dashed line marks the 70% global test accuracy threshold; dotted vertical lines mark the first round reaching it—30 vs. 60 rounds at K = 20 and 59 vs. 100 at K = 100. 0 5k 10k 15k 20k 25k Total generated samples 55 60 65 70 75 Accuracy (%) FedAvg (no generation) =4 =8 =12 =16 =20 FedEAS Uniform (a) Accuracy at matched budgets 0 10 20 30 F… view at source ↗
Figure 5
Figure 5. Figure 5: Budget sensitivity on CIFAR-10 (α = 0.1, K = 20) under matched FedEAS– Uniform budgets. (a) Last-20-round accuracy as the total generation budget grows. (b) Selection of the default budget parameter. The budget path traces the imbalance reduction 1− I(B)/I(0) against the synthetic fraction of the training data as β grows; markers are the sweep values. The dashed lines mark the two selection conditions. Onl… view at source ↗
Figure 1
Figure 1. Figure 1: Closed-form normalized heterogeneity residual h(r) = ((1 − r)/(1 + r))2 from Theorem 1. The residual decreases rapidly at small generation ratios and flattens to￾ward full balance (r = 1), illustrating the diminishing marginal reduction from addi￾tional generation. B.1 Extension to K Clients and C Classes (Sketch) Continuing the closed-form perspective above, we sketch how the qualitative principle can ext… view at source ↗
Figure 2
Figure 2. Figure 2: Uncurated random samples from the class-conditional DDPM (native 32×32, shown enlarged with nearest-neighbor interpolation for visibility) used by all DDPM￾based augmentation policies. Rows correspond to the ten CIFAR-10 classes. Samples are provided for transparency; this work focuses on allocation rather than per-sample fidelity, and generator-quality variance is orthogonal to the allocation policy (Sect… view at source ↗

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

64 extracted references · 64 canonical work pages · 7 internal anchors

  1. [1]

    Artificial intelligence and statistics , pages=

    Communication-efficient learning of deep networks from decentralized data , author=. Artificial intelligence and statistics , pages=. 2017 , organization=

    work page 2017

  2. [2]

    Federated Learning with Non-IID Data

    Federated learning with non-iid data , author=. arXiv preprint arXiv:1806.00582 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    Proceedings of Machine learning and systems , volume=

    Federated optimization in heterogeneous networks , author=. Proceedings of Machine learning and systems , volume=

    work page

  4. [4]

    International conference on machine learning , pages=

    Scaffold: Stochastic controlled averaging for federated learning , author=. International conference on machine learning , pages=. 2020 , organization=

    work page 2020

  5. [5]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Model-contrastive federated learning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  6. [6]

    International conference on machine learning , pages=

    Data-free knowledge distillation for heterogeneous federated learning , author=. International conference on machine learning , pages=. 2021 , organization=

    work page 2021

  7. [7]

    Stable Diffusion-based Data Augmentation for Federated Learning with Non-IID Data

    Stable diffusion-based data augmentation for federated learning with non-iid data , author=. arXiv preprint arXiv:2405.07925 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

    work page

  9. [9]

    The 11th International Conference on Learning Representations (ICLR 2023)(01/05/2023-05/05/2023, Kigali, Rwanda) , year=

    Is synthetic data from generative models ready for image recognition? , author=. The 11th International Conference on Learning Representations (ICLR 2023)(01/05/2023-05/05/2023, Kigali, Rwanda) , year=

    work page 2023

  10. [10]

    Transactions on Machine Learning Research , year=

    Synthetic Data from Diffusion Models Improves ImageNet Classification , author=. Transactions on Machine Learning Research , year=

    work page

  11. [11]

    Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining , pages=

    Fedrs: Federated learning with restricted softmax for label distribution non-iid data , author=. Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining , pages=

    work page

  12. [12]

    International Conference on Machine Learning , pages=

    Federated learning with label distribution skew via logits calibration , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    work page 2022

  13. [13]

    International Conference on Learning Representations , year=

    On the Convergence of FedAvg on Non-IID Data , author=. International Conference on Learning Representations , year=

    work page

  14. [14]

    International Conference on Learning Representations , year=

    FedMix: Approximation of Mixup under Mean Augmented Federated Learning , author=. International Conference on Learning Representations , year=

    work page

  15. [15]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    A simple data augmentation for feature distribution skewed federated learning , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    work page

  16. [16]

    International Conference on Learning Representations , volume=

    Effective data augmentation with diffusion models , author=. International Conference on Learning Representations , volume=

    work page

  17. [17]

    Advances in Neural Information Processing Systems , volume=

    No fear of heterogeneity: Classifier calibration for federated learning with non-iid data , author=. Advances in Neural Information Processing Systems , volume=

    work page

  18. [18]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    No fear of classifier biases: Neural collapse inspired federated learning with synthetic and fixed classifier , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

    work page

  19. [19]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Exploiting label skews in federated learning with model concatenation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    work page

  20. [20]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Exploring vacant classes in label-skewed federated learning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    work page

  21. [21]

    Tianfei Zhou and Ender Konukoglu , booktitle=. Fed. 2023 , url=

    work page 2023

  22. [22]

    Advances in Neural Information Processing Systems , volume=

    Fedavp: Augment local data via shared policy in federated learning , author=. Advances in Neural Information Processing Systems , volume=

    work page

  23. [23]

    Differentially Private Diffusion Models Generate Useful Synthetic Images

    Differentially private diffusion models generate useful synthetic images , author=. arXiv preprint arXiv:2302.13861 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [24]

    Foundations and trends in machine learning , volume=

    Advances and open problems in federated learning , author=. Foundations and trends in machine learning , volume=. 2021 , publisher=

    work page 2021

  25. [25]

    Measuring the Effects of Non-Identical Data Distribution for Federated Visual Classification

    Measuring the effects of non-identical data distribution for federated visual classification , author=. arXiv preprint arXiv:1909.06335 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 1909

  26. [26]

    Transactions on Machine Learning Research , issn=

    On the Unreasonable Effectiveness of Federated Averaging with Heterogeneous Data , author=. Transactions on Machine Learning Research , issn=. 2024 , url=

    work page 2024

  27. [27]

    2022 IEEE 38th international conference on data engineering (ICDE) , pages=

    Federated learning on non-iid data silos: An experimental study , author=. 2022 IEEE 38th international conference on data engineering (ICDE) , pages=. 2022 , organization=

    work page 2022

  28. [28]

    International Conference on Learning Representations , year=

    Federated Learning Based on Dynamic Regularization , author=. International Conference on Learning Representations , year=

    work page

  29. [29]

    Advances in neural information processing systems , volume=

    Tackling the objective inconsistency problem in heterogeneous federated optimization , author=. Advances in neural information processing systems , volume=

    work page

  30. [30]

    International Conference on Learning Representations , year=

    mixup: Beyond Empirical Risk Minimization , author=. International Conference on Learning Representations , year=

    work page

  31. [31]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Cutmix: Regularization strategy to train strong classifiers with localizable features , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

    work page

  32. [32]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  33. [33]

    International Conference on Learning Representations , year=

    Denoising Diffusion Implicit Models , author=. International Conference on Learning Representations , year=

    work page

  34. [34]

    Advances in neural information processing systems , volume=

    Diffusion models beat gans on image synthesis , author=. Advances in neural information processing systems , volume=

    work page

  35. [35]

    NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications , year=

    Classifier-Free Diffusion Guidance , author=. NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications , year=

    work page 2021

  36. [36]

    International Conference on Learning Representations , year=

    Adaptive Federated Optimization , author=. International Conference on Learning Representations , year=

    work page

  37. [37]

    IEEE signal processing magazine , volume=

    Federated learning: Challenges, methods, and future directions , author=. IEEE signal processing magazine , volume=. 2020 , publisher=

    work page 2020

  38. [38]

    International conference on machine learning , pages=

    Improved denoising diffusion probabilistic models , author=. International conference on machine learning , pages=. 2021 , organization=

    work page 2021

  39. [39]

    International Conference on Learning Representations , year=

    Score-Based Generative Modeling through Stochastic Differential Equations , author=. International Conference on Learning Representations , year=

    work page

  40. [40]

    Communications of the ACM , volume=

    Generative adversarial networks , author=. Communications of the ACM , volume=. 2020 , publisher=

    work page 2020

  41. [41]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Analyzing and improving the image quality of stylegan , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  42. [42]

    European Conference on Computer Vision , pages=

    Adversarial diffusion distillation , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    work page 2024

  43. [43]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Deep residual learning for image recognition , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    work page

  44. [44]

    2009 , institution=

    Learning multiple layers of features from tiny images , author=. 2009 , institution=

    work page 2009

  45. [45]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Autoaugment: Learning augmentation strategies from data , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  46. [46]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops , pages=

    Randaugment: Practical automated data augmentation with a reduced search space , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops , pages=

    work page

  47. [47]

    Improved Regularization of Convolutional Neural Networks with Cutout

    Improved regularization of convolutional neural networks with cutout , author=. arXiv preprint arXiv:1708.04552 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  48. [48]

    Journal of big data , volume=

    A survey on image data augmentation for deep learning , author=. Journal of big data , volume=. 2019 , publisher=

    work page 2019

  49. [49]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Fake it till you make it: Learning transferable representations from synthetic imagenet clones , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  50. [50]

    Advances in Neural Information Processing Systems , volume=

    Stablerep: Synthetic images from text-to-image models make strong visual representation learners , author=. Advances in Neural Information Processing Systems , volume=

    work page

  51. [51]

    Advances in neural information processing systems , volume=

    Ensemble distillation for robust model fusion in federated learning , author=. Advances in neural information processing systems , volume=

    work page

  52. [52]

    International conference on machine learning , pages=

    Ditto: Fair and robust federated learning through personalization , author=. International conference on machine learning , pages=. 2021 , organization=

    work page 2021

  53. [53]

    International conference on machine learning , pages=

    Exploiting shared representations for personalized federated learning , author=. International conference on machine learning , pages=. 2021 , organization=

    work page 2021

  54. [54]

    Personalized Federated Learning: A Meta-Learning Approach

    Personalized federated learning: A meta-learning approach , author=. arXiv preprint arXiv:2002.07948 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 2002

  55. [55]

    Advances in neural information processing systems , volume=

    Learning imbalanced datasets with label-distribution-aware margin loss , author=. Advances in neural information processing systems , volume=

    work page

  56. [56]

    International Conference on Learning Representations , year=

    Decoupling Representation and Classifier for Long-Tailed Recognition , author=. International Conference on Learning Representations , year=

    work page

  57. [57]

    Proceedings of the 2016 ACM SIGSAC conference on computer and communications security , pages=

    Deep learning with differential privacy , author=. Proceedings of the 2016 ACM SIGSAC conference on computer and communications security , pages=

    work page 2016

  58. [58]

    International Conference on Learning Representations , year=

    Learning Differentially Private Recurrent Language Models , author=. International Conference on Learning Representations , year=

    work page

  59. [59]

    International workshop on trustworthy federated learning , pages=

    Federated learning with gan-based data synthesis for non-iid clients , author=. International workshop on trustworthy federated learning , pages=. 2022 , organization=

    work page 2022

  60. [60]

    Chinese Conference on Pattern Recognition and Computer Vision (PRCV) , pages=

    Federated learning based on diffusion model to cope with non-iid data , author=. Chinese Conference on Pattern Recognition and Computer Vision (PRCV) , pages=. 2023 , organization=

    work page 2023

  61. [61]

    IEEE Access , volume=

    FedDrip: Federated learning with diffusion-generated synthetic image , author=. IEEE Access , volume=. 2025 , publisher=

    work page 2025

  62. [62]

    IEEE Transactions on Mobile Computing , volume=

    Feature matching data synthesis for non-IID federated learning , author=. IEEE Transactions on Mobile Computing , volume=. 2024 , publisher=

    work page 2024

  63. [63]

    Federated Learning Empowered by Generative Content

    Federated learning empowered by generative content , author=. arXiv preprint arXiv:2312.05807 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  64. [64]

    IEEE Transactions on Mobile Computing , volume=

    Filling the missing: Exploring generative AI for enhanced federated learning over heterogeneous mobile edge devices , author=. IEEE Transactions on Mobile Computing , volume=. 2024 , publisher=

    work page 2024