Data-Free Contribution Estimation in Federated Learning using Gradient von Neumann Entropy

Asim Ukaye; Karthik Nandakumar; Mubarak Abdu-Aguye; Nurbek Tastan

arxiv: 2604.22562 · v1 · submitted 2026-04-24 · 💻 cs.LG · cs.AI· cs.CV· cs.DC

Data-Free Contribution Estimation in Federated Learning using Gradient von Neumann Entropy

Asim Ukaye , Mubarak Abdu-Aguye , Nurbek Tastan , Karthik Nandakumar This is my paper

Pith reviewed 2026-05-08 12:23 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CVcs.DC

keywords federated learningclient contribution estimationvon Neumann entropydata-free methodsnon-IID datagradient updates

0 comments

The pith

Matrix von Neumann entropy of final-layer gradients estimates client contributions in federated learning without validation data.

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

The paper introduces a data-free method to gauge each client's value in federated learning by computing the spectral entropy of its final-layer gradient updates. This signal captures the diversity of information a client brings and requires neither server-side validation sets nor any client metadata. Experiments on CIFAR-10/100, FEMNIST, and FedISIC under varied non-IID splits show that the resulting scores track closely with the accuracy a client would obtain if trained in isolation. The approach is instantiated in two schemes, SpectralFed and SpectralFuse, that use the entropy either directly for weighting or fused with alignment information for stability.

Core claim

The matrix von Neumann entropy of the final-layer updates serves as a reliable indicator of client contribution, achieving high correlation with standalone client accuracies on CIFAR-10/100, FEMNIST, and FedISIC under various non-IID regimes when used in SpectralFed and SpectralFuse schemes.

What carries the argument

Matrix von Neumann entropy of the final-layer gradient updates, which quantifies the diversity of information contributed by each client.

Load-bearing premise

The matrix von Neumann entropy of the final-layer updates specifically and reliably measures the diversity and usefulness of the information contributed by each client, rather than being dominated by model architecture, optimization noise, or other training dynamics.

What would settle it

A new benchmark or architecture where entropy scores show consistently low or negative correlation with standalone client accuracies would falsify the central claim.

Figures

Figures reproduced from arXiv: 2604.22562 by Asim Ukaye, Karthik Nandakumar, Mubarak Abdu-Aguye, Nurbek Tastan.

**Figure 1.** Figure 1: Illustration of the proposed SpectralFuse approach. We incorporate the von Neumann Entropy and Class-specific Shapley values view at source ↗

**Figure 2.** Figure 2: Top left: Three clients with equal amount of data but varying number of labels from the MNIST dataset. Bottom left: Eigenvalue spectrum corresponding to the final-layer updates of each client. Top right: Accuracies obtained by each client on a hold out test set as training progresses. Bottom right: Spectral entropy for each client as the training progresses. von Neumann entropy of \rho is given as S(\rho )… view at source ↗

**Figure 3.** Figure 3: Schematic of the Rank Adaptive Kalman Filter used to view at source ↗

**Figure 4.** Figure 4: SpectralFed (green) shows consistent Pearson correla view at source ↗

**Figure 5.** Figure 5: The RAKF’s working principle demonstrated for a run view at source ↗

**Figure 6.** Figure 6: Effect of choice of filter process noise covariance view at source ↗

**Figure 8.** Figure 8: Sensitivity of choice of Q and ϵ on the average Pearson correlation of the fused weights and standalone accuracies for different splits on the CIFAR-10 dataset. Average values of correlation are robust with respect to changes in filter hyperparameters. results in a generally higher value of R view at source ↗

**Figure 9.** Figure 9: Free-rider detection accuracy of the proposed methods view at source ↗

**Figure 10.** Figure 10: Average Pearson correlation of the entropy of the gradient for each layer with the standalone accuracy. The top row displays view at source ↗

read the original abstract

Client contribution estimation in Federated Learning is necessary for identifying clients' importance and for providing fair rewards. Current methods often rely on server-side validation data or self-reported client information, which can compromise privacy or be susceptible to manipulation. We introduce a data-free signal based on the matrix von Neumann (spectral) entropy of the final-layer updates, which measures the diversity of the information contributed. We instantiate two practical schemes: (i) SpectralFed, which uses normalized entropy as aggregation weights, and (ii) SpectralFuse, which fuses entropy with class-specific alignment via a rank-adaptive Kalman filter for per-round stability. Across CIFAR-10/100 and the naturally partitioned FEMNIST and FedISIC benchmarks, entropy-derived scores show a consistently high correlation with standalone client accuracy under diverse non-IID regimes - without validation data or client metadata. We compare our results with data-free contribution estimation baselines and show that spectral entropy serves as a useful indicator of client contribution.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Entropy on final-layer gradients gives a clean data-free signal, but validating it only against standalone accuracy leaves the contribution claim short.

read the letter

The paper's main idea is to score clients by the matrix von Neumann entropy of their final-layer gradient updates. This produces a number that reflects information diversity without any server-side validation set or client metadata. They turn the score into two aggregation rules: SpectralFed just normalizes the entropy into weights, while SpectralFuse blends it with class alignment through a rank-adaptive Kalman filter for round-to-round stability. Both are tested on CIFAR-10/100 and the naturally partitioned FEMNIST and FedISIC sets under varied non-IID splits, and the entropy numbers line up well with how accurately each client could train by itself. That correlation is the concrete result they report, and it is new relative to the data-free baselines they cite. The approach is simple enough that someone could re-implement the entropy step quickly and check it on their own data. The limitation is that standalone accuracy is not the same as a client's marginal effect on the global model. In federated training, value depends on how a client's distribution fills gaps left by everyone else; training the client alone ignores those interactions and can be driven mainly by local data volume. The abstract gives no leave-one-out retraining numbers or Shapley-style comparisons, so the reported correlations do not directly confirm that the entropy signal improves aggregation fairness or accuracy. The summary also omits error bars, ablation on the entropy matrix size, and full tables, which makes it hard to judge how sensitive the numbers are to the Kalman parameters or to different layer choices. This work is aimed at people building incentive or auditing layers for production federated systems that cannot move data. It is worth sending to referees because the data-free entropy angle is distinct and the benchmarks are standard, even though the validation needs to be tightened against actual marginal contribution measures.

Referee Report

3 major / 2 minor

Summary. The paper proposes a data-free approach to client contribution estimation in federated learning based on the matrix von Neumann (spectral) entropy of final-layer gradient updates, which is intended to capture the diversity of information contributed by each client. Two instantiations are presented: SpectralFed, which directly uses normalized entropy values as aggregation weights, and SpectralFuse, which fuses the entropy signal with class-specific alignment scores via a rank-adaptive Kalman filter. Experiments on CIFAR-10/100 and naturally partitioned FEMNIST/FedISIC under varied non-IID regimes report consistently high correlations between the entropy-derived scores and each client's standalone accuracy (i.e., accuracy of a model trained only on its local data), together with comparisons against other data-free baselines.

Significance. A reliable data-free signal for contribution estimation would be valuable for privacy-preserving incentive mechanisms in federated learning. The method's complete avoidance of server-side validation data or client metadata is a clear strength, and the use of spectral entropy provides an interpretable, architecture-agnostic measure of gradient diversity. If the reported correlations hold under more rigorous validation against marginal (rather than standalone) contribution, the approach could influence practical FL deployments; however, the current evidence base leaves the link between entropy and actual federated utility incompletely established.

major comments (3)

[Abstract] Abstract: the central claim that von Neumann entropy of final-layer updates measures client contribution rests on reported high correlations with standalone client accuracy. Standalone accuracy, however, ignores inter-client complementarity and can be driven primarily by local data volume or class balance; no leave-one-out retraining, Shapley-value, or other marginal-contribution ground truth is provided to test whether the entropy signal predicts the client's effect on the global model after aggregation.
[Abstract] Abstract / Experimental results: the abstract states 'consistently high correlation' across CIFAR-10/100, FEMNIST, and FedISIC but supplies no numerical coefficients, confidence intervals, p-values, or full per-client tables. Without these details or an ablation on the entropy computation itself (e.g., choice of final layer, matrix size, or numerical stability), the robustness of the correlation claim cannot be assessed.
[SpectralFuse] SpectralFuse description: the rank-adaptive Kalman filter introduces several tunable fusion and rank-adaptation parameters. The manuscript should report an ablation or sensitivity analysis showing how performance and stability vary with these hyperparameters; their selection could indirectly influence the stability results attributed to the entropy signal.

minor comments (2)

[Methods] Notation: the precise definition of the matrix von Neumann entropy (including any normalization, trace operations, or handling of singular values) should be stated explicitly in the methods section with a numbered equation.
[Figures] Figure clarity: correlation plots should include error bars or per-seed variability to allow visual assessment of consistency across non-IID regimes.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback on our work. We address each major comment below, agreeing where the manuscript can be strengthened and providing clarifications on our design choices. Revisions will be made to improve transparency and rigor.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim that von Neumann entropy of final-layer updates measures client contribution rests on reported high correlations with standalone client accuracy. Standalone accuracy, however, ignores inter-client complementarity and can be driven primarily by local data volume or class balance; no leave-one-out retraining, Shapley-value, or other marginal-contribution ground truth is provided to test whether the entropy signal predicts the client's effect on the global model after aggregation.

Authors: We agree that standalone accuracy is an imperfect proxy for true marginal contribution, since it does not explicitly quantify complementarity or the incremental effect on the aggregated global model. Our use of this metric was driven by computational practicality: exact marginal measures such as Shapley values or full leave-one-out retraining scale poorly with the number of clients and rounds in realistic FL benchmarks. The consistently high correlations we observe nevertheless indicate that spectral entropy captures data diversity relevant to local utility. In revision we will add an explicit limitations paragraph in the introduction and experiments sections discussing this distinction, referencing related marginal-contribution literature, and reporting a limited leave-one-out study on CIFAR-10 to provide supplementary evidence. revision: yes
Referee: [Abstract] Abstract / Experimental results: the abstract states 'consistently high correlation' across CIFAR-10/100, FEMNIST, and FedISIC but supplies no numerical coefficients, confidence intervals, p-values, or full per-client tables. Without these details or an ablation on the entropy computation itself (e.g., choice of final layer, matrix size, or numerical stability), the robustness of the correlation claim cannot be assessed.

Authors: The full manuscript contains per-client correlation tables and numerical values in Section 4, yet the abstract indeed omits specific coefficients and statistical details. We will revise the abstract to report representative Pearson and Spearman coefficients together with 95% confidence intervals for the primary benchmarks. We will also add a dedicated ablation subsection examining sensitivity of the entropy signal to final-layer selection, gradient matrix size, and numerical stability under different normalization schemes. revision: yes
Referee: [SpectralFuse] SpectralFuse description: the rank-adaptive Kalman filter introduces several tunable fusion and rank-adaptation parameters. The manuscript should report an ablation or sensitivity analysis showing how performance and stability vary with these hyperparameters; their selection could indirectly influence the stability results attributed to the entropy signal.

Authors: We concur that the tunable parameters of the rank-adaptive Kalman filter warrant explicit sensitivity analysis. The revised manuscript will include a new subsection reporting how contribution-score stability and downstream accuracy change when varying the fusion weight, rank-adaptation threshold, and process-noise covariance. This analysis will help separate the contribution of the entropy signal from the fusion mechanism. revision: yes

Circularity Check

0 steps flagged

No significant circularity; entropy signal computed independently of validation targets

full rationale

The paper defines the matrix von Neumann entropy directly from final-layer gradient updates and uses it to produce contribution scores and aggregation weights (SpectralFed, SpectralFuse). This computation does not fit parameters to standalone client accuracy or any other target metric; the reported correlations are presented as post-hoc empirical observations rather than definitional or fitted outcomes. The rank-adaptive Kalman filter introduces fusion parameters for stability, but these are not shown to be tuned against the accuracy correlations that support the main claim. No load-bearing self-citations, uniqueness theorems, or ansatzes reduce the central derivation to its own inputs. The chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The method rests on the domain assumption that spectral entropy of final-layer gradients quantifies contribution diversity; the Kalman filter component likely introduces free parameters for fusion and rank adaptation.

free parameters (1)

Kalman filter fusion and rank-adaptation parameters
SpectralFuse fuses entropy with class-specific alignment inside a rank-adaptive Kalman filter; such filters require tuning parameters whose values are not specified in the abstract.

axioms (1)

domain assumption The matrix von Neumann entropy of final-layer gradient updates measures the diversity of information contributed by a client.
This assumption underpins both SpectralFed and SpectralFuse and is invoked to justify the data-free claim.

pith-pipeline@v0.9.0 · 5482 in / 1270 out tokens · 43017 ms · 2026-05-08T12:23:28.722176+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages

[1]

Com- putation of the von neumann entropy of large matrices via trace estimators and rational krylov methods.Numerische Mathematik, 155(3):377–414, 2023

Michele Benzi, Michele Rinelli, and Igor Simunec. Com- putation of the von neumann entropy of large matrices via trace estimators and rational krylov methods.Numerische Mathematik, 155(3):377–414, 2023. 1, 3

work page 2023
[2]

Leaf: A benchmark for federated settings,

Sebastian Caldas, Sai Meher Karthik Duddu, Peter Wu, Tian Li, Jakub Kone ˇcn`y, H Brendan McMahan, Virginia Smith, and Ameet Talwalkar. Leaf: A benchmark for federated set- tings.arXiv preprint arXiv:1812.01097, 2018. 2

work page arXiv 2018
[3]

Fair federated medical image segmentation via client contri- bution estimation

Meirui Jiang, Holger R Roth, Wenqi Li, Dong Yang, Can Zhao, Vishwesh Nath, Daguang Xu, Qi Dou, and Ziyue Xu. Fair federated medical image segmentation via client contri- bution estimation. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 16302–16311, 2023. 1, 2

work page 2023
[4]

Fantastic generalization mea- sures and where to find them

Yiding Jiang, Behnam Neyshabur, Hossein Mobahi, Dilip Krishnan, and Samy Bengio. Fantastic generalization mea- sures and where to find them. InInternational Conference on Learning Representations, 2020. 3

work page 2020
[5]

Advances and open problems in federated learning.Foundations and trends in machine learning, 14(1-2):1–210, 2021

Peter Kairouz and H Brendan McMahan. Advances and open problems in federated learning.Foundations and trends in machine learning, 14(1-2):1–210, 2021. 1

work page 2021
[6]

A new approach to linear filtering and prediction problems.Transactions of the ASME–Journal of Basic Engineering, 82(Series D):35–45, 1960

Rudolph Emil Kalman. A new approach to linear filtering and prediction problems.Transactions of the ASME–Journal of Basic Engineering, 82(Series D):35–45, 1960. 1, 4

work page 1960
[7]

Data valuation and detections in federated learning

Wenqian Li, Shuran Fu, Fengrui Zhang, and Yan Pang. Data valuation and detections in federated learning. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12027–12036, 2024. 3

work page 2024
[8]

Collabo- rative fairness in federated learning

Lingjuan Lyu, Xinyi Xu, Qian Wang, and Han Yu. Collabo- rative fairness in federated learning. InFederated Learning: Privacy and Incentive, pages 189–204. Springer, 2020. 1, 2

work page 2020
[9]

Guiding neural collapse: Optimising towards the nearest simplex equiangular tight frame.Advances in Neural Information Processing Systems, 37:35544–35573, 2024

Evan Markou, Thalaiyasingam Ajanthan, and Stephen Gould. Guiding neural collapse: Optimising towards the nearest simplex equiangular tight frame.Advances in Neural Information Processing Systems, 37:35544–35573, 2024. 4

work page 2024
[10]

Implicit self- regularization in deep neural networks: Evidence from ran- dom matrix theory and implications for learning.Journal of Machine Learning Research, 22(165):1–73, 2021

Charles H Martin and Michael W Mahoney. Implicit self- regularization in deep neural networks: Evidence from ran- dom matrix theory and implications for learning.Journal of Machine Learning Research, 22(165):1–73, 2021. 3

work page 2021
[11]

Martin, Tongsu, Peng, and Michael W

Charles H. Martin, Tongsu, Peng, and Michael W. Mahoney. Predicting trends in the quality of state-of-the-art neural net- works without access to training or testing data.Nature Com- munications, 12(1):4122, 2021. 3, 5, 6

work page 2021
[12]

Communication- Efficient Learning of Deep Networks from Decentralized Data

Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication- Efficient Learning of Deep Networks from Decentralized Data. InProceedings of the 20th International Conference on Artificial Intelligence and Statistics, pages 1273–1282. PMLR, 2017. 1, 2, 3, 6

work page 2017
[13]

FLamby: Datasets and Benchmarks for Cross-Silo Federated Learning in Realistic Healthcare Settings.Advances in Neural Information Pro- cessing Systems, 35:5315–5334, 2022

Jean Ogier du Terrail, Samy-Safwan Ayed, Edwige Cyf- fers, Felix Grimberg, Chaoyang He, Regis Loeb, Paul Man- gold, Tanguy Marchand, Othmane Marfoq, Erum Mush- taq, Boris Muzellec, Constantin Philippenko, Santiago Silva, Maria Tele ´nczuk, Shadi Albarqouni, Salman Avestimehr, Aur´elien Bellet, Aymeric Dieuleveut, Martin Jaggi, Sai Pra- neeth Karimireddy, ...

work page 2022
[14]

Vardan Papyan, X. Y . Han, and David L. Donoho. Prevalence of neural collapse during the terminal phase of deep learning training.Proceedings of the National Academy of Sciences, 117(40):24652–24663, 2020. 4

work page 2020
[15]

Springer, 2002

Asher Peres.Quantum theory: concepts and methods. Springer, 2002. 1

work page 2002
[16]

Reddi, Zachary Charles, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Kone ˇcn´y, Sanjiv Kumar, and Hugh Brendan McMahan

Sashank J. Reddi, Zachary Charles, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Kone ˇcn´y, Sanjiv Kumar, and Hugh Brendan McMahan. Adaptive Federated Optimiza- tion. InInternational Conference on Learning Representa- tions, 2021. 3

work page 2021
[17]

Redefining Contribu- tions: Shapley-Driven Federated Learning

Nurbek Tastan, Samar Fares, Toluwani Aremu, Samuel Horv´ath, and Karthik Nandakumar. Redefining Contribu- tions: Shapley-Driven Federated Learning. InProceedings of the Thirty-Third International Joint Conference on Artifi- cial Intelligence, IJCAI-24, pages 5009–5017. International Joint Conferences on Artificial Intelligence Organization,

work page
[18]

1, 2, 3, 5, 6

Main Track. 1, 2, 3, 5, 6

work page
[19]

Aequa: Fair Model Rewards in Collaborative Learning via Slimmable Networks

Nurbek Tastan, Samuel Horv ´ath, and Karthik Nandakumar. Aequa: Fair Model Rewards in Collaborative Learning via Slimmable Networks. InProceedings of the 42nd Inter- national Conference on Machine Learning, pages 59210– 59236. PMLR, 2025

work page 2025
[20]

CYCle: Choosing Your Collaborators Wisely to Enhance Collaborative Fairness in Decentralized Learning.Transac- tions on Machine Learning Research, 2025

Nurbek Tastan, Samuel Horv ´ath, and Karthik Nandakumar. CYCle: Choosing Your Collaborators Wisely to Enhance Collaborative Fairness in Decentralized Learning.Transac- tions on Machine Learning Research, 2025. 2

work page 2025
[21]

Fedpca: Noise-robust fair federated learning via performance-capacity analysis.arXiv preprint arXiv:2503.10567, 2025

Nannan Wu, Zengqiang Yan, Nong Sang, Li Yu, and Chang Wen Chen. Fedpca: Noise-robust fair federated learning via performance-capacity analysis.arXiv preprint arXiv:2503.10567, 2025. 3

work page arXiv 2025
[22]

Gradient Driven Rewards to Guarantee Fairness in Collaborative Ma- chine Learning

Xinyi Xu, Lingjuan Lyu, Xingjun Ma, Chenglin Miao, Chuan Sheng Foo, and Bryan Kian Hsiang Low. Gradient Driven Rewards to Guarantee Fairness in Collaborative Ma- chine Learning. InAdvances in Neural Information Process- ing Systems, pages 16104–16117. Curran Associates, Inc.,

work page
[23]

A new perspective to boost performance fairness for medical federated learning

Yunlu Yan, Lei Zhu, Yuexiang Li, Xinxing Xu, Rick Siow Mong Goh, Yong Liu, Salman Khan, and Chun-Mei Feng. A new perspective to boost performance fairness for medical federated learning. InInternational Conference on Medical Image Computing and Computer-Assisted Interven- tion, pages 13–23. Springer, 2024. 3

work page 2024
[24]

Gonzalez, Kannan Ramchandran, Charles H

Yaoqing Yang, Ryan Theisen, Liam Hodgkinson, Joseph E. Gonzalez, Kannan Ramchandran, Charles H. Martin, and Michael W. Mahoney. Evaluating natural language process- ing models with generalization metrics that do not need ac- cess to any training or testing data, 2023. arXiv:2202.02842 [cs]. 3 9

work page arXiv 2023
[25]

A Geometric Analysis of Neural Collapse with Unconstrained Features

Zhihui Zhu, Tianyu Ding, Jinxin Zhou, Xiao Li, Chong You, Jeremias Sulam, and Qing Qu. A Geometric Analysis of Neural Collapse with Unconstrained Features. InAdvances in Neural Information Processing Systems, pages 29820– 29834. Curran Associates, Inc., 2021. 4 10 Data-Free Contribution Estimation in Federated Learning using Gradient von Neumann Entropy S...

work page 2021

[1] [1]

Com- putation of the von neumann entropy of large matrices via trace estimators and rational krylov methods.Numerische Mathematik, 155(3):377–414, 2023

Michele Benzi, Michele Rinelli, and Igor Simunec. Com- putation of the von neumann entropy of large matrices via trace estimators and rational krylov methods.Numerische Mathematik, 155(3):377–414, 2023. 1, 3

work page 2023

[2] [2]

Leaf: A benchmark for federated settings,

Sebastian Caldas, Sai Meher Karthik Duddu, Peter Wu, Tian Li, Jakub Kone ˇcn`y, H Brendan McMahan, Virginia Smith, and Ameet Talwalkar. Leaf: A benchmark for federated set- tings.arXiv preprint arXiv:1812.01097, 2018. 2

work page arXiv 2018

[3] [3]

Fair federated medical image segmentation via client contri- bution estimation

Meirui Jiang, Holger R Roth, Wenqi Li, Dong Yang, Can Zhao, Vishwesh Nath, Daguang Xu, Qi Dou, and Ziyue Xu. Fair federated medical image segmentation via client contri- bution estimation. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 16302–16311, 2023. 1, 2

work page 2023

[4] [4]

Fantastic generalization mea- sures and where to find them

Yiding Jiang, Behnam Neyshabur, Hossein Mobahi, Dilip Krishnan, and Samy Bengio. Fantastic generalization mea- sures and where to find them. InInternational Conference on Learning Representations, 2020. 3

work page 2020

[5] [5]

Advances and open problems in federated learning.Foundations and trends in machine learning, 14(1-2):1–210, 2021

Peter Kairouz and H Brendan McMahan. Advances and open problems in federated learning.Foundations and trends in machine learning, 14(1-2):1–210, 2021. 1

work page 2021

[6] [6]

A new approach to linear filtering and prediction problems.Transactions of the ASME–Journal of Basic Engineering, 82(Series D):35–45, 1960

Rudolph Emil Kalman. A new approach to linear filtering and prediction problems.Transactions of the ASME–Journal of Basic Engineering, 82(Series D):35–45, 1960. 1, 4

work page 1960

[7] [7]

Data valuation and detections in federated learning

Wenqian Li, Shuran Fu, Fengrui Zhang, and Yan Pang. Data valuation and detections in federated learning. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12027–12036, 2024. 3

work page 2024

[8] [8]

Collabo- rative fairness in federated learning

Lingjuan Lyu, Xinyi Xu, Qian Wang, and Han Yu. Collabo- rative fairness in federated learning. InFederated Learning: Privacy and Incentive, pages 189–204. Springer, 2020. 1, 2

work page 2020

[9] [9]

Guiding neural collapse: Optimising towards the nearest simplex equiangular tight frame.Advances in Neural Information Processing Systems, 37:35544–35573, 2024

Evan Markou, Thalaiyasingam Ajanthan, and Stephen Gould. Guiding neural collapse: Optimising towards the nearest simplex equiangular tight frame.Advances in Neural Information Processing Systems, 37:35544–35573, 2024. 4

work page 2024

[10] [10]

Implicit self- regularization in deep neural networks: Evidence from ran- dom matrix theory and implications for learning.Journal of Machine Learning Research, 22(165):1–73, 2021

Charles H Martin and Michael W Mahoney. Implicit self- regularization in deep neural networks: Evidence from ran- dom matrix theory and implications for learning.Journal of Machine Learning Research, 22(165):1–73, 2021. 3

work page 2021

[11] [11]

Martin, Tongsu, Peng, and Michael W

Charles H. Martin, Tongsu, Peng, and Michael W. Mahoney. Predicting trends in the quality of state-of-the-art neural net- works without access to training or testing data.Nature Com- munications, 12(1):4122, 2021. 3, 5, 6

work page 2021

[12] [12]

Communication- Efficient Learning of Deep Networks from Decentralized Data

Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication- Efficient Learning of Deep Networks from Decentralized Data. InProceedings of the 20th International Conference on Artificial Intelligence and Statistics, pages 1273–1282. PMLR, 2017. 1, 2, 3, 6

work page 2017

[13] [13]

FLamby: Datasets and Benchmarks for Cross-Silo Federated Learning in Realistic Healthcare Settings.Advances in Neural Information Pro- cessing Systems, 35:5315–5334, 2022

Jean Ogier du Terrail, Samy-Safwan Ayed, Edwige Cyf- fers, Felix Grimberg, Chaoyang He, Regis Loeb, Paul Man- gold, Tanguy Marchand, Othmane Marfoq, Erum Mush- taq, Boris Muzellec, Constantin Philippenko, Santiago Silva, Maria Tele ´nczuk, Shadi Albarqouni, Salman Avestimehr, Aur´elien Bellet, Aymeric Dieuleveut, Martin Jaggi, Sai Pra- neeth Karimireddy, ...

work page 2022

[14] [14]

Vardan Papyan, X. Y . Han, and David L. Donoho. Prevalence of neural collapse during the terminal phase of deep learning training.Proceedings of the National Academy of Sciences, 117(40):24652–24663, 2020. 4

work page 2020

[15] [15]

Springer, 2002

Asher Peres.Quantum theory: concepts and methods. Springer, 2002. 1

work page 2002

[16] [16]

Reddi, Zachary Charles, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Kone ˇcn´y, Sanjiv Kumar, and Hugh Brendan McMahan

Sashank J. Reddi, Zachary Charles, Manzil Zaheer, Zachary Garrett, Keith Rush, Jakub Kone ˇcn´y, Sanjiv Kumar, and Hugh Brendan McMahan. Adaptive Federated Optimiza- tion. InInternational Conference on Learning Representa- tions, 2021. 3

work page 2021

[17] [17]

Redefining Contribu- tions: Shapley-Driven Federated Learning

Nurbek Tastan, Samar Fares, Toluwani Aremu, Samuel Horv´ath, and Karthik Nandakumar. Redefining Contribu- tions: Shapley-Driven Federated Learning. InProceedings of the Thirty-Third International Joint Conference on Artifi- cial Intelligence, IJCAI-24, pages 5009–5017. International Joint Conferences on Artificial Intelligence Organization,

work page

[18] [18]

1, 2, 3, 5, 6

Main Track. 1, 2, 3, 5, 6

work page

[19] [19]

Aequa: Fair Model Rewards in Collaborative Learning via Slimmable Networks

Nurbek Tastan, Samuel Horv ´ath, and Karthik Nandakumar. Aequa: Fair Model Rewards in Collaborative Learning via Slimmable Networks. InProceedings of the 42nd Inter- national Conference on Machine Learning, pages 59210– 59236. PMLR, 2025

work page 2025

[20] [20]

CYCle: Choosing Your Collaborators Wisely to Enhance Collaborative Fairness in Decentralized Learning.Transac- tions on Machine Learning Research, 2025

Nurbek Tastan, Samuel Horv ´ath, and Karthik Nandakumar. CYCle: Choosing Your Collaborators Wisely to Enhance Collaborative Fairness in Decentralized Learning.Transac- tions on Machine Learning Research, 2025. 2

work page 2025

[21] [21]

Fedpca: Noise-robust fair federated learning via performance-capacity analysis.arXiv preprint arXiv:2503.10567, 2025

Nannan Wu, Zengqiang Yan, Nong Sang, Li Yu, and Chang Wen Chen. Fedpca: Noise-robust fair federated learning via performance-capacity analysis.arXiv preprint arXiv:2503.10567, 2025. 3

work page arXiv 2025

[22] [22]

Gradient Driven Rewards to Guarantee Fairness in Collaborative Ma- chine Learning

Xinyi Xu, Lingjuan Lyu, Xingjun Ma, Chenglin Miao, Chuan Sheng Foo, and Bryan Kian Hsiang Low. Gradient Driven Rewards to Guarantee Fairness in Collaborative Ma- chine Learning. InAdvances in Neural Information Process- ing Systems, pages 16104–16117. Curran Associates, Inc.,

work page

[23] [23]

A new perspective to boost performance fairness for medical federated learning

Yunlu Yan, Lei Zhu, Yuexiang Li, Xinxing Xu, Rick Siow Mong Goh, Yong Liu, Salman Khan, and Chun-Mei Feng. A new perspective to boost performance fairness for medical federated learning. InInternational Conference on Medical Image Computing and Computer-Assisted Interven- tion, pages 13–23. Springer, 2024. 3

work page 2024

[24] [24]

Gonzalez, Kannan Ramchandran, Charles H

Yaoqing Yang, Ryan Theisen, Liam Hodgkinson, Joseph E. Gonzalez, Kannan Ramchandran, Charles H. Martin, and Michael W. Mahoney. Evaluating natural language process- ing models with generalization metrics that do not need ac- cess to any training or testing data, 2023. arXiv:2202.02842 [cs]. 3 9

work page arXiv 2023

[25] [25]

A Geometric Analysis of Neural Collapse with Unconstrained Features

Zhihui Zhu, Tianyu Ding, Jinxin Zhou, Xiao Li, Chong You, Jeremias Sulam, and Qing Qu. A Geometric Analysis of Neural Collapse with Unconstrained Features. InAdvances in Neural Information Processing Systems, pages 29820– 29834. Curran Associates, Inc., 2021. 4 10 Data-Free Contribution Estimation in Federated Learning using Gradient von Neumann Entropy S...

work page 2021