Data Overvaluation Attack and Truthful Data Valuation in Federated Learning

Chuan Xiao; Jianbin Qin; Makoto Onizuka; Masatoshi Yoshikawa; Shuyuan Zheng; Sudong Cai; Yang Cao

REVIEW 3 major objections 4 minor 1 cited by

Data Overvaluation Attack and Truthful Data Valuation in Federated Learning

T0 review · 3 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Strategic clients can inflate their data value in federated learning; the paper proposes Truth-Shapley as the unique truthful valuation metric.

desk verdict The data overvaluation attack is real and worth knowing, but Truth-Shapley's uniqueness theorem is false as stated, so the paper needs major revision before the theory can be trusted. read the letter →

arxiv 2502.00494 v3 pith:D4P5Y6FR submitted 2025-02-01 cs.CR cs.AIcs.LG

Shuyuan Zheng , Sudong Cai , Chuan Xiao , Yang Cao , Jianbin Qin , Masatoshi Yoshikawa , Makoto Onizuka This is my paper · ORCID

classification cs.CRcs.AIcs.LG MSC 91A1291A2668T05

keywords datavaluationfederatedlearningovervaluationattackTruth-ShapleyBayesianincentivecompatibilityShapleyvaluestrategicreportinglinearmetrics

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

This paper argues that in federated learning, a strategic client can inflate its own data value—and deflate everyone else's—by misreporting its local data during the model retraining steps that data valuation requires. It introduces the data overvaluation attack, which works against every linear valuation metric, including Shapley value, leave-one-out, Beta Shapley, and Banzhaf; experiments show attacker value rising by 77% on average for Shapley and 97% for LOO. The paper then proposes Truth-Shapley, a linear, efficient valuation that combines client-level and block-level Shapley-style averaging over full client datasets. Truth-Shapley is claimed to be the unique metric satisfying the six valuation axioms while being Bayesian incentive compatible, meaning that under the paper's key assumption, each client's best strategy is to report its true data. If correct, this turns data valuation from a process that rewards manipulation into one where truthful reporting is the dominant strategy.

What carries the argument

The load-bearing object is the coefficient function $\beta_i(S)$ from Lemma 3.1: linearity forces each client's value to be a weighted sum of subset utilities, and the attack is a rule for flipping the sign of each $v(S)$ according to the signs of $\beta_i(S)$ and $\beta_{-i}(S)$. The truthfulness result rests on Bayesian incentive compatibility (Definition 4.1) plus Assumption 4.2, which says each client believes its own full dataset $D_i$ maximizes expected utility; under that belief, only metrics whose $\beta_i(S)$ vanish for every misreported subset, and are nonnegative for subsets equal to the client's full dataset, can be truthful (Theorem 4.3). Truth-Shapley is built to meet this characterization: it evaluates client-level value by Shapley averaging over utilities of full-dataset coalitions, and block-level value by an inner Shapley averaging inside each client, so the two-level construction carries both efficiency and incentive compatibility.

What would settle it

Construct a client whose held-back dataset strictly improves expected model utility over its reported subset for some subset $S$, run the paper's Algorithm 1 with that client, and check whether its Truth-Shapley value changes: if it changes, the claimed Bayesian incentive compatibility fails. More directly, exhibit a utility function $v$ and beliefs $\sigma$ satisfying the paper's setup but with $\mathbb{E}[v(D_i \cup D_{-i})] < \mathbb{E}[v(\hat D_i \cup D_{-i})]$ for some misreport $\hat D_i$; then Assumption 4.2 is violated and the Theorem 4.5 guarantee no longer applies.

Watch

Extended reading notes

Core claim

The central discovery is that any linear data valuation metric can be written as $\varphi_i(D_N, v) = \sum_{S \subseteq D_N} \beta_i(S)\, v(S)$, so a client who knows the signs of $\beta_i(S)$ can selectively raise or lower the retrained model utilities to increase its own value and reduce others'. The data overvaluation attack does exactly this by positive or negative data augmentation during subset-model retraining. Against this, the paper characterizes the linear, efficient, Bayesian incentive compatible metrics: the value of client $i$ may depend only on utilities $v(D_C)$ of coalitions of clients' full datasets, with nonnegative coefficients. Truth-Shapley is the instance that applies Shapley averaging at the client level and then again inside each client's blocks; Theorem 4.5 states it is Bayesian incentive compatible, and Theorem 4.6 states it is the unique metric satisfying efficiency, linearity, and the four adapted dummy/symmetry axioms.

Load-bearing premise

The whole truthfulness guarantee rests on Assumption 4.2: every client must believe, based on its own information, that its full local dataset is the best dataset it could submit for the joint model, so that no misreport can look better to it.

Editorial extensions

If this is right

Under the data overvaluation attack, Shapley value, leave-one-out, Beta Shapley, and Banzhaf value produce inflated attacker values and distorted selection, while Truth-Shapley shows a 0.0% change in the paper's experiments.
Truth-Shapley stays at or near best data-selection utility even without attacks, so its robustness does not come at the cost of selection quality.
Any linear metric that fails the $\beta_i(S)$ characterization of Theorem 4.3 is vulnerable in principle, giving a check for future valuation designs.
The two-level Shapley construction inherits existing Shapley acceleration techniques, so practical computation is feasible despite the exponential worst-case retraining count.
If Assumption 4.2 holds, truthful reporting is an equilibrium: no client can raise its expected value or lower others' by misreporting.

Reading between the lines

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

The guarantee is subjective: it depends on each client's belief, so a client who knows it holds back higher-quality data, or a poisoning attacker, breaks the premise; the paper itself flags this in Appendix C.
The attack assumes clients can estimate or infer the signs of $\beta_i(S)$ and $\beta_{-i}(S)$; a server that hides the subset identity or randomizes valuation subsets may blunt the attack, though the paper shows estimation suffices.
In vertical federated learning, positive augmentation is hard because feature spaces differ, so the attack and Truth-Shapley may transfer only partially, as Appendix C notes.
A testable extension is to measure how much Truth-Shapley degrades as clients' beliefs become misspecified, or when a client's 'best' dataset is not the union of its blocks.

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.

Desk Editor's Note

read the letter

Hi,

The data overvaluation attack is the real contribution here: a clear demonstration that a client can inflate its Shapley/LOO value by misreporting local data during subset retraining, with solid experimental backing across several FL algorithms. The BIC characterization (Theorems 4.3–4.4) is also plausible and gives useful structure to the space of attack-resistant linear metrics. I'd point anyone working on strategic behavior in federated data valuation at this paper.

The headline theoretical claim, Theorem 4.6, is false as stated. Truth-Shapley does not satisfy the paper's own DUM-IB axiom. Simple counterexample: two clients, client 1 has blocks {a,b}, client 2 has {c}. Set v(empty)=0, v({a})=1, v({b})=0, v({a,b})=1, v({c})=2, v({a,c})=2, v({b,c})=2, v({a,b,c})=2. This game is monotone. DUM-IB holds: block a has marginal 1 over all subsets of client 1's other blocks; block b has marginal 0; block c has marginal 2. Now compute TSV: client-level values are 0.5 and 1.5. The internal utility on client 1's blocks gives block values 0.5 for a and 0 for b; client 2's block gets 1.5. So φ(a)=0.5 ≠ v({a})=1, and φ(c)=1.5 ≠ 2. DUM-IB is violated. The Appendix D proof applies Shapley uniqueness separately at the client and block levels, but the axioms couple: block-level Shapley uses an internal utility that depends on other clients' blocks, so dummy behavior in the inner sense is not preserved.

This is not a minor typo; the uniqueness theorem is the paper's main advertised contribution. The BIC result (Theorem 4.5) may still be salvageable, and the attack remains valid. But as written, the theoretical centerpiece doesn't hold up. I'll give the authors credit for flagging Assumption 4.2 and dual-objective attackers in Appendix C.

Who should read this: people studying data valuation attacks, not people looking for a provably truthful metric. I'd send it to a serious referee because the attack is novel and the flaw is identifiable and possibly repairable, but I'd expect major revision before publication.

Referee Report

3 major / 4 minor

Summary. The paper proposes the data overvaluation attack, in which a strategic client in federated learning misreports local representations during subset-model retraining to inflate its value under any linear data valuation metric. It then studies Bayesian incentive compatible (BIC) data valuation, gives a characterization of linear BIC metrics under Assumption 4.2, and introduces Truth-Shapley (TSV), claimed to be the unique metric satisfying EFF, LIN, DUM-C, DUM-IB, SYM-C, and SYM-IB. Experiments across several datasets and FL algorithms show that the attack inflates SV, LOO, BSV, and BV values while leaving TSV unchanged.

Significance. The attack formulation is well motivated, and the experimental comparison is unusually thorough: it covers multiple FL algorithms, reports standard errors, and ablates utility and reward functions. The BIC characterization is an interesting axiomatic exercise that could be useful beyond the specific metric. However, the paper's central theoretical claims are not correct: the uniqueness theorem for Truth-Shapley is false, and the stated characterization theorem is too strong. Since these theorems are the advertised main contributions, the paper cannot be accepted in its current form.

major comments (3)

[§4.2, Theorem 4.6 and Appendix D] Truth-Shapley does not satisfy DUM-IB, so the uniqueness theorem is false. Consider N={1,2}, D1={x,y}, D2={z}, and let v(∅)=0, v({x})=1, v({y})=0, v({x,y})=1, v({z})=2, v({x,z})=v({y,z})=v({x,y,z})=2. DUM-IB holds for x (check S=∅ and S={y}), for y, and for z, so the axiom requires φ_{1,x}=1, φ_{1,y}=0, φ_{2,z}=2. Computing TSV gives client-level values φ1=0.5 and φ2=1.5, and block-level values φ_{1,x}=0.5, φ_{1,y}=0, φ_{2,z}=1.5. The reason is that the inner utility w(S)=φ1(D−1∪S,v) is not the restriction of v to D1: dummy behavior of a block in v is not preserved under the client-level Shapley transform. The proof in Appendix D applies Shapley uniqueness separately at client and block level, but the axioms are coupled through w(S), so the composed metric does not inherit DUM-IB.
[§4.1, Theorem 4.3] The stated iff condition is not necessary for BIC because it forces β_i(S)=0 also for subsets S that contain no block of client i, even though client i has no report over such S and cannot influence v(S). For example, with N={1,2}, D1={a}, D2={b}, the TSV client-level value is φ1=0.5v({a})+0.5v({a,b})−0.5v({b})−0.5v(∅), so β1({b})=−0.5 while D1^{{b}}=∅≠D1. Theorem 4.3 would therefore declare TSV non-BIC, contradicting Theorem 4.5. The theorem and its proof must be restricted to subsets S with i∈N(S); as printed, the characterization is incorrect.
[Appendix D, proof of Theorem 4.5] The displayed formula for β_i^{TSV}(D_C) has the signs reversed: for C containing i, the coefficient of v(D_C) in φ_i^{TSV} is +wSV(C\{i}|N), and for C not containing i it is −wSV(C|N). The printed formula would make β_i(D_C)<0 for i∈C, which is the opposite of what Theorem 4.4 requires. The statement of Theorem 4.5 can be recovered with corrected signs, but the proof as written is internally inconsistent.

minor comments (4)

[§4.2] The notation D_C (the set of datasets of clients in C) and D_C (the union of data blocks of clients in C) is distinguished only in words; the paper should use distinct symbols throughout.
[Appendix D, proof of Theorem 4.6] The axioms EFF-C and EFF-IB are introduced informally inside the proof but are not stated among the six axioms in Theorem 4.6; they should be defined in the theorem statement.
[Table 1] The TSV columns are exactly 0.0 by construction, because TSV's client-level coefficients vanish on all partial block subsets; the table would be more informative if it also reported the block-level distortions, which are nonzero even in the simple counterexample above.
[Eq. (1) and Algorithm 1] The notation S+ and \hat S is used without a consistent formal definition of S+ when S is a set of data blocks; it should be introduced once at first use.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the BIC characterization and Truth-Shapley's robustness are derived from an explicit assumption and the linearity structure, not from fitted parameters or self-citation; claimed theoretical flaws are soundness issues, not circularity.

full rationale

The paper's derivation chain is self-contained rather than circular. Lemma 3.1 derives the linear functional representation of any linear valuation metric from basis games; Lemma 3.3 proves the attack's effect from the attack's own sign conditions; Theorems 4.3 and 4.4 characterize BIC under Assumption 4.2, which is an explicit axiom about each client's subjective beliefs, not a restatement of the target theorem. Truth-Shapley is then constructed to satisfy the characterization, so Theorem 4.5 and the observed 0.0 attack impact in Tables 1-2 follow by proof rather than by fitting. The experiments on TSV are consistency checks of a theorem, not predictions obtained from fitted inputs. The only self-citations ([8], [46], [47]) appear in Appendix C in the context of privacy-preserving FL and differential privacy, and none is load-bearing for the attack, BIC characterization, or TSV construction. Appendix C itself flags the core limitation: 'Assumption 4.2 is the core assumption of this paper, which implicitly assumes that the attacker's dataset Di is not a poisoning dataset.' This is an honest scope limitation, not a circular step. The claimed uniqueness theorem (Theorem 4.6) is proven by two separate applications of Shapley's external uniqueness theorem (Theorem A.1, [33]) to the client-level and block-level games; even if that proof is mathematically insufficient because the axioms are coupled across levels, that would be a soundness defect, not circularity, because the conclusion is not assumed as an input. Accordingly, the circularity score is low and no circular step is catalogued.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

No free parameters or invented entities. The results rest on Assumption 4.2 and standard cooperative game theory; the central flaw is not an extra parameter but a mismatch between stated axioms and the proof.

assumptions (3)

domain assumption Assumption 4.2 (Subjectively Optimal Data): for any S with D_i^S = D_i and any \hat D_i^S, E[v(D_i \cup \hat D_{-i}^S)] >= E[v(\hat D_i^S \cup \hat D_{-i}^S)].
This is the core behavioral assumption that makes truthful reporting optimal; if it fails (e.g., poisoning attackers), BIC is not guaranteed. Acknowledged in Appendix C.
standard math Shapley uniqueness theorem (Theorem A.1): the Shapley value is the unique valuation satisfying LIN, EFF, DUM, and SYM.
Used to prove uniqueness of TSV at client and block levels; standard cooperative game theory.
standard math Any linear valuation can be written as a weighted sum of subset utilities (Lemma 3.1).
Basis game decomposition of linear functionals on the space of utility functions; used throughout the characterization.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Data Overvaluation Attack and Truthful Data Valuation in Federated Learning." pith.science (2026). https://pith.science/paper/D4P5Y6FR

@misc{pith2026250200494,
  author       = {Pith},
  title        = {Pith review of: Data Overvaluation Attack and Truthful Data Valuation in Federated Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D4P5Y6FR}},
  note         = {Machine review of arXiv:2502.00494}
}

read the original abstract

In collaborative machine learning (CML), data valuation, i.e., evaluating the contribution of each client's data to the machine learning model, has become a critical task for incentivizing and selecting positive data contributions. However, existing studies often assume that clients engage in data valuation truthfully, overlooking the practical motivation for clients to exaggerate their contributions. To unlock this threat, this paper introduces the data overvaluation attack, enabling strategic clients to have their data significantly overvalued in federated learning, a widely adopted paradigm for decentralized CML. Furthermore, we propose a Bayesian truthful data valuation metric, named Truth-Shapley. Truth-Shapley is the unique metric that guarantees some promising axioms for data valuation while ensuring that clients' optimal strategy is to perform truthful data valuation under certain conditions. Our experiments demonstrate the vulnerability of existing data valuation metrics to the proposed attack and validate the robustness and effectiveness of Truth-Shapley.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Quotient Semivalues for False-Name-Resistant Data Attribution
cs.GT 2026-05 unverdicted novelty 8.0 of 10

Quotient semivalues enable false-name-proof data attribution by clustering identities with a canonical representative and prove that exact Shapley fairness is incompatible with unrestricted false-name resistance on mo...

Reference graph

Works this paper leans on

47 extracted references · 41 canonical work pages · cited by 1 Pith paper

[1]

UCI Machine Learning Repo sitory, 2019

Apartment for Rent Classiﬁed. UCI Machine Learning Repo sitory, 2019. DOI: https://doi.org/10.24432/C5X623

work page doi:10.24432/c5x623 2019
[2]

A marke tplace for data: An algorithmic solution

Anish Agarwal, Munther Dahleh, and Tuhin Sarkar. A marke tplace for data: An algorithmic solution. In Proceedings of the 2019 ACM Conference on Economics and Comp utation, pages 701–726, 2019

work page 2019
[3]

Truthful dat a acquisition via peer prediction

Yiling Chen, Yiheng Shen, and Shuran Zheng. Truthful dat a acquisition via peer prediction. Advances in Neural Information Processing Systems , 33:18194–18204, 2020

work page 2020
[4]

Feature sele ction based on the shapley value

Shay B Cohen, Gideon Dror, and Eytan Ruppin. Feature sele ction based on the shapley value. In Proceedings of IJCAI, pages 1–6, 2005

work page 2005
[5]

Detection of inﬂuential observation in li near regression

R Dennis Cook. Detection of inﬂuential observation in li near regression. T echnometrics, 19 (1):15–18, 1977

work page 1977
[6]

Bayesian ince ntive compatible beliefs

Claude d’Aspremont and L-A Gérard-V aret. Bayesian ince ntive compatible beliefs. Journal of Mathematical Economics , 10(1):83–103, 1982

work page 1982
[7]

Calibrating noise to sen- sitivity in private data analysis

Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Sm ith. Calibrating noise to sen- sitivity in private data analysis. In Theory of Cryptography: Third Theory of Cryptography Conference, TCC 2006, New Y ork, NY , USA, March 4-7, 2006. Proceedings 3, pages 265–284. Springer, 2006

work page 2006
[8]

Privacy-enhanced d atabase synthesis for bench- mark publishing

Y unqing Ge, Jianbin Qin, Shuyuan Zheng, Y ongrui Zhong, B o Tang, Y u-Xuan Qiu, Rui Mao, Y e Y uan, Makoto Onizuka, and Chuan Xiao. Privacy-enhanced d atabase synthesis for bench- mark publishing. Proceedings of the VLDB Endowment , 18(2):413–425, 2024

work page 2024

Show all 47 references

[9]

Data shapley: Equitable valuation of data for machine learning

Amirata Ghorbani and James Zou. Data shapley: Equitable valuation of data for machine learning. In International conference on machine learning , pages 2242–2251. PMLR, 2019

2019
[10]

Efﬁcient task-speci ﬁc data valuation for nearest neighbor algorithms

Ruoxi Jia, David Dao, Boxin Wang, Frances Ann Hubis, Nez ihe Merve Gurel, Bo Li, Ce Zhang, Costas Spanos, and Dawn Song. Efﬁcient task-speci ﬁc data valuation for nearest neighbor algorithms. Proceedings of the VLDB Endowment , 12(11):1610–1623, 2019

2019
[11]

Towards efﬁci ent data valuation based on the shapley value

Ruoxi Jia, David Dao, Boxin Wang, Frances Ann Hubis, Nic k Hynes, Nezihe Merve Gürel, Bo Li, Ce Zhang, Dawn Song, and Costas J Spanos. Towards efﬁci ent data valuation based on the shapley value. In The 22nd International Conference on Artiﬁcial Intelligen ce and Statistics, p...

2019
[12]

LA V A: Data valuation without pre-speciﬁed learn ing algorithms

Hoang Anh Just, Feiyang Kang, Tianhao Wang, Yi Zeng, Mye ongseob Ko, Ming Jin, and Ruoxi Jia. LA V A: Data valuation without pre-speciﬁed learn ing algorithms. In The Eleventh International Conference on Learning Representations , 2023

2023
[13]

Understanding black-box p redictions via inﬂuence functions

Pang Wei Koh and Percy Liang. Understanding black-box p redictions via inﬂuence functions. In International conference on machine learning , pages 1885–1894. PMLR, 2017

2017
[14]

Learning mult iple layers of features from tiny images

Alex Krizhevsky, Geoffrey Hinton, et al. Learning mult iple layers of features from tiny images. 2009. 10

2009
[15]

Beta shapley: a uniﬁed and n oise-reduced data valuation framework for machine learning

Y ongchan Kwon and James Zou. Beta shapley: a uniﬁed and n oise-reduced data valuation framework for machine learning. In International Conference on Artiﬁcial Intelligence and Statistics, pages 8780–8802. PMLR, 2022

2022
[16]

Efﬁcient c omputation and analysis of dis- tributional shapley values

Y ongchan Kwon, Manuel A Rivas, and James Zou. Efﬁcient c omputation and analysis of dis- tributional shapley values. In International Conference on Artiﬁcial Intelligence and St atistics, pages 793–801. PMLR, 2021

2021
[17]

Gradient-based learning applied to document recognition

Y ann LeCun, Léon Bottou, Y oshua Bengio, and Patrick Haf fner. Gradient-based learning applied to document recognition. Proceedings of the IEEE , 86(11):2278–2324, 1998

1998
[18]

Fedmd: Heterogenous federat ed learning via model distillation

Daliang Li and Junpu Wang. Fedmd: Heterogenous federat ed learning via model distillation. arXiv preprint arXiv:1910.03581 , 2019

1910 arXiv
[19]

Measuring the effect of training data on deep learning predi ctions via randomized experiments

Jinkun Lin, Anqi Zhang, Mathias Lécuyer, Jinyang Li, Au rojit Panda, and Siddhartha Sen. Measuring the effect of training data on deep learning predi ctions via randomized experiments. In International Conference on Machine Learning , pages 13468–13504. PMLR, 2022

2022
[20]

Dis- tributionally robust data valuation

Xiaoqiang Lin, Xinyi Xu, Zhaoxuan Wu, See-Kiong Ng, and Bryan Kian Hsiang Low. Dis- tributionally robust data valuation. In Proceedings of the 41st International Conference on Machine Learning, 2024

2024
[21]

On shapley v alue in data assemblage under independent utility

Xuan Luo, Jian Pei, Zicun Cong, and Cheng Xu. On shapley v alue in data assemblage under independent utility. Proceedings of the VLDB Endowment , 15(11):2761–2773, 2022

2022
[22]

Fast shapley value com- putation in data assemblage tasks as cooperative simple gam es

Xuan Luo, Jian Pei, Cheng Xu, Wenjie Zhang, and Jianlian g Xu. Fast shapley value com- putation in data assemblage tasks as cooperative simple gam es. Proceedings of the ACM on Management of Data, 2(1):1–28, 2024

2024
[23]

Communication-efﬁcient learning of deep networks fro m decentralized data

Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hamp son, and Blaise Aguera y Ar- cas. Communication-efﬁcient learning of deep networks fro m decentralized data. In Artiﬁcial intelligence and statistics, pages 1273–1282, 2017

2017
[24]

Cortez, and P

Sérgio Moro, P . Cortez, and P . Rita. Bank Marketing. UCIMachine Learning Repository, 2014. DOI: https://doi.org/10.24432/C5K306

2014 doi
[25]

Game of gra dients: Mitigating irrelevant clients in federated learning

Lokesh Nagalapatti and Ramasuri Narayanam. Game of gra dients: Mitigating irrelevant clients in federated learning. In Proceedings of the AAAI Conference on Artiﬁcial Intelligen ce, vol- ume 35, pages 9046–9054, 2021

2021
[26]

Trade-off between pay- off and model rewards in shapley-fair collaborative machin e learning

Quoc Phong Nguyen, Bryan Kian Hsiang Low, and Patrick Ja illet. Trade-off between pay- off and model rewards in shapley-fair collaborative machin e learning. Advances in Neural Information Processing Systems, 35:30542–30553, 2022

2022
[27]

Data valuatio n without training of a model

Ki Nohyun, Hoyong Choi, and Hye Won Chung. Data valuatio n without training of a model. In The Eleventh International Conference on Learning Represe ntations, 2022

2022
[28]

Collaborative machine learning markets with data-replication-robust payments

Olga Ohrimenko, Shruti Tople, and Sebastian Tschiatsc hek. Collaborative machine learning markets with data-replication-robust payments. arXiv preprint arXiv:1911.09052 , 2019

1911 arXiv
[29]

Karl Pearson. Liii. on lines and planes of closest ﬁt to s ystems of points in space. The London, Edinburgh, and Dublin philosophical magazine and journal o f science, 2(11):559–572, 1901

1901
[30]

Estimating training data inﬂuence by tracing gradient descent

Garima Pruthi, Frederick Liu, Satyen Kale, and Mukund S undararajan. Estimating training data inﬂuence by tracing gradient descent. Advances in Neural Information Processing Sys- tems, 33:19920–19930, 2020

2020
[31]

Sentence-bert: Sente nce embeddings using siamese bert-networks

Nils Reimers and Iryna Gurevych. Sentence-bert: Sente nce embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in N at- ural Language Processing . Association for Computational Linguistics, 11 2019. URL https://arxiv.org/abs/1908.10084

2019 arXiv
[32]

Cs-shap ley: class-wise shapley values for data valuation in classiﬁcation

Stephanie Schoch, Haifeng Xu, and Y angfeng Ji. Cs-shap ley: class-wise shapley values for data valuation in classiﬁcation. Advances in Neural Information Processing Systems , 35: 34574–34585, 2022. 11

2022
[33]

A value for n-person games

Lloyd S Shapley. A value for n-person games. Contribution to the Theory of Games , 2, 1953

1953
[34]

Proﬁt alloc ation for federated learning

Tianshu Song, Y ongxin Tong, and Shuyue Wei. Proﬁt alloc ation for federated learning. In 2019 IEEE International Conference on Big Data (Big Data) , pages 2577–2586. IEEE, 2019

2019
[35]

Splitfed: When federated learning meets split learning

Chandra Thapa, Pathum Chamikara Mahawaga Arachchige, Seyit Camtepe, and Lichao Sun. Splitfed: When federated learning meets split learning. In Proceedings of the AAAI conference on artiﬁcial intelligence , volume 36, pages 8485–8493, 2022

2022
[36]

Derdava: Deletion- robust data valuation for machine learning

Xiao Tian, Rachael Hwee Ling Sim, Jue Fan, and Bryan Kian Hsiang Low. Derdava: Deletion- robust data valuation for machine learning. In Proceedings of the AAAI Conference on Artiﬁcial Intelligence, volume 38, pages 15373–15381, 2024

2024
[37]

Data banzhaf: A robust data valuation framework for machine learning

Jiachen T Wang and Ruoxi Jia. Data banzhaf: A robust data valuation framework for machine learning. In International Conference on Artiﬁcial Intelligence and St atistics, pages 6388–
[38]

A privacy- friendly approach to data valuation

Jiachen Tianhao Wang, Y uqing Zhu, Y u-Xiang Wang, RuoxiJia, and Prateek Mittal. A privacy- friendly approach to data valuation. Advances in Neural Information Processing Systems , 36, 2024

2024
[39]

Helpful or harmful data? ﬁne-tuning-free shapley attribut ion for explaining language model predictions

Jingtan Wang, Xiaoqiang Lin, Rui Qiao, Chuan-Sheng Foo , and Bryan Kian Hsiang Low. Helpful or harmful data? ﬁne-tuning-free shapley attribut ion for explaining language model predictions. In Proceedings of the 41st International Conference on Machin e Learning, 2024

2024
[40]

Davinz: Data valuation using deep neural networks at initialization

Zhaoxuan Wu, Y ao Shu, and Bryan Kian Hsiang Low. Davinz: Data valuation using deep neural networks at initialization. In International Conference on Machine Learning , pages 24150–24176. PMLR, 2022

2022
[41]

Equitable data valuation meets the right to be forgotten in model marke ts

Haocheng Xia, Jinfei Liu, Jian Lou, Zhan Qin, Kui Ren, Y a ng Cao, and Li Xiong. Equitable data valuation meets the right to be forgotten in model marke ts. Proceedings of the VLDB Endowment, 16(11):3349–3362, 2023

2023
[42]

P-shapley: Shapley values on probabilistic classiﬁers

Haocheng Xia, Xiang Li, Junyuan Pang, Jinfei Liu, Kui Re n, and Li Xiong. P-shapley: Shapley values on probabilistic classiﬁers. Proceedings of the VLDB Endowment , 17(7):1737–1750, 2024

2024
[43]

Fashion-m nist: a novel image dataset for bench- marking machine learning algorithms

Han Xiao, Kashif Rasul, and Roland V ollgraf. Fashion-m nist: a novel image dataset for bench- marking machine learning algorithms. arXiv preprint arXiv:1708.07747 , 2017

2017 arXiv
[44]

V alidation free and replication robust volume-based data valuation

Xinyi Xu, Zhaoxuan Wu, Chuan Sheng Foo, and Bryan Kian Hs iang Low. V alidation free and replication robust volume-based data valuation. Advances in Neural Information Processing Systems, 34:10837–10848, 2021

2021
[45]

Model shapley: eq- uitable model valuation with black-box access

Xinyi Xu, Thanh Lam, Chuan Sheng Foo, and Bryan Kian Hsia ng Low. Model shapley: eq- uitable model valuation with black-box access. Advances in Neural Information Processing Systems, 36, 2024

2024
[46]

Fl-market: Trading private models in federated learning

Shuyuan Zheng, Y ang Cao, Masatoshi Y oshikawa, Huizhon g Li, and Qiang Y an. Fl-market: Trading private models in federated learning. In 2022 IEEE International Conference on Big Data (Big Data) , pages 1525–1534. IEEE, 2022

2022
[47]

job" and the attribute

Shuyuan Zheng, Y ang Cao, and Masatoshi Y oshikawa. Secu re shapley value for cross-silo federated learning. Proceedings of the VLDB Endowment , 16(7):1657–1670, 2023. 12 Broader Impact For the industry, this paper identiﬁes a new attack method th at poses a trust crisis for d...

2023

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

UCI Machine Learning Repo sitory, 2019

Apartment for Rent Classiﬁed. UCI Machine Learning Repo sitory, 2019. DOI: https://doi.org/10.24432/C5X623

work page doi:10.24432/c5x623 2019

[2] [2]

A marke tplace for data: An algorithmic solution

Anish Agarwal, Munther Dahleh, and Tuhin Sarkar. A marke tplace for data: An algorithmic solution. In Proceedings of the 2019 ACM Conference on Economics and Comp utation, pages 701–726, 2019

work page 2019

[3] [3]

Truthful dat a acquisition via peer prediction

Yiling Chen, Yiheng Shen, and Shuran Zheng. Truthful dat a acquisition via peer prediction. Advances in Neural Information Processing Systems , 33:18194–18204, 2020

work page 2020

[4] [4]

Feature sele ction based on the shapley value

Shay B Cohen, Gideon Dror, and Eytan Ruppin. Feature sele ction based on the shapley value. In Proceedings of IJCAI, pages 1–6, 2005

work page 2005

[5] [5]

Detection of inﬂuential observation in li near regression

R Dennis Cook. Detection of inﬂuential observation in li near regression. T echnometrics, 19 (1):15–18, 1977

work page 1977

[6] [6]

Bayesian ince ntive compatible beliefs

Claude d’Aspremont and L-A Gérard-V aret. Bayesian ince ntive compatible beliefs. Journal of Mathematical Economics , 10(1):83–103, 1982

work page 1982

[7] [7]

Calibrating noise to sen- sitivity in private data analysis

Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Sm ith. Calibrating noise to sen- sitivity in private data analysis. In Theory of Cryptography: Third Theory of Cryptography Conference, TCC 2006, New Y ork, NY , USA, March 4-7, 2006. Proceedings 3, pages 265–284. Springer, 2006

work page 2006

[8] [8]

Privacy-enhanced d atabase synthesis for bench- mark publishing

Y unqing Ge, Jianbin Qin, Shuyuan Zheng, Y ongrui Zhong, B o Tang, Y u-Xuan Qiu, Rui Mao, Y e Y uan, Makoto Onizuka, and Chuan Xiao. Privacy-enhanced d atabase synthesis for bench- mark publishing. Proceedings of the VLDB Endowment , 18(2):413–425, 2024

work page 2024