acceptodds
Under review as a conference paper at ICLR 2027

Shapley Identification under Partial Information in Federated Learning

Abstract

In privacy-preserving federated learning, clients can evaluate two utilities locally without revealing their updates: the utility of their update alone and the utility lost when they are removed from the full aggregate. A scorer can collect these measurements across clients. We study how much this collection determines the full vector of Shapley values and develop a theory of what can be inferred from these measurements. In one completed federated round, their midpoint recovers Shapley whenever interactions involve at most two clients. With three-client interactions, however, the same observations can correspond to different Shapley values. For games with nonnegative effects involving at most three clients, we characterize each client’s interval of compatible Shapley values and the rankings shared by all compatible Shapley vectors. Beyond the pairwise case, smoothness still controls the error: the midpoint differs from the true Shapley value only at cubic order in the update size, and no rule using the same measurements can improve this order. Given bounds on curvature and update size, this approximation yields certified score intervals. In our experiments, these intervals contain every exact Shapley value and determine nearly all pairwise client rankings.

Primary area:
interpretability and explainable AI
Keywords:
federated learning, Shapley values, contribution valuation, client attribution, endpoint measurements, identifiability, approximation bounds, certified intervals, cooperative game theory

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_9LrcOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether Shapley values can be accurately recovered in privacy-preserving federated learning when only two measurements are available for each client: its standalone gain and removal loss. The results show that Shapley values can be exactly recovered when interactions are limited to pairs of clients. However, when higher-order interactions are present, Shapley values may no longer be uniquely identified.

Critical strengths

S1. The research question is underexplored and meaningful for improving the efficiency and security of contribution evaluation in FL.

S2. The research question is answered with a complete set of theoretical characterizations.

S3. The presentation is good, and the claims are carefully stated.

Critical weaknesses

W1. The key contribution Theorem 6 gives exact intervals only when the true game has nonnegative effects through order three. However, only 2.2% of the empirical observations satisfy this assumption.

W2. The scope of the experiments is quite limited. Only three datasets and single-round FedAvg are considered, which limits the generalizability of the empirical findings.

Questions for authors

Q1. How can we calculate the L value in Theorem 7 for neural networks?

4 comments

Thank you for taking the time to read and review our work. We have uploaded an updated version of the paper. The theorem numbering and all section, table and appendix references below refer to that version. Original Theorems 6–8 are now Theorems 5–7.

The replies below distinguish the structural identified set from the operational smooth certificate, report the added comparisons, and give the neural derivative calculation.

W1. The structural theorem gives exact intervals only when the true game has nonnegative effects through order three, while the reported 2.2% figure suggests limited empirical applicability.

The structural theorem characterizes information left unresolved by exact recovery. In Figure 1, two positive third-order games share si=1s_i=1, mi=7m_i=7 and V=14V=14 but have different top contributors. Theorem 5 identifies every compatible Shapley vector in the class; for that example each coordinate's sharp interval is [10/3,4][10/3,4]. It determines the remaining uncertainty, including the rankings common to all compatible vectors.

The 2.2% figure measures endpoint compatibility with this class, not true-game membership. All six compatible observations in the diagnostic have true Shapley vectors outside at least one structural interval. Section 3.3 therefore presents the result as a structural benchmark requiring independent membership justification, and Appendix C.6 states what the compatibility test establishes.

The experiments use the separate smooth certificate, which permits signed interactions of arbitrary order under public curvature and update controls. The structural compatibility rate does not assess that certificate's validity. Sections 5.3–5.4 instead measure its informative decisions and retained progress, as reported under W2.

W2. The original experiments cover only three datasets and single-round FedAvg, limiting their empirical generality.

We addressed operating-range generality with a paired sensitivity on both logistic datasets and added a learned-representation task on MNIST. The sensitivity reuses each profile across clip/scale choices, so the comparison measures the controls' effect with the underlying updates fixed. At full scale, the highlighted settings retain 87.6% and 86.3% median unclipped round gain and certify 482/560 and 471/560 pairs, with 14/20 and 13/20 winners. Section 5.3, Table 1, Figure 2 and Appendix C.2 give the full comparison and separate coordinate-only from combined certificates.

The added MNIST task evaluates a validation-only selection rule and certification for head updates over a fixed learned encoder. The original coordinate rule certifies 526/560 test pairs, with at least 23/28 in every replicate and median 27/28. Applying the combined rule at the same selected settings gives 533/560 pairs and 18/20 winners. Section 5.4 and Appendix C.3 report the original study, and Appendix C.7 reports the fixed-selection refinement. The partition replicates share image pools; these counts describe the evaluated completed rounds.

The formal results hold over their stated game classes; the experiments test whether their certificates give useful decisions in the evaluated rounds. The eight-client design permits exhaustive coalition evaluation to audit the intervals against exact Shapley, while the scorer still receives only endpoints and public controls. The sensitivity and learned-representation study establish the measured operating range and its progress cost. Section 5 and Appendix C now make that theory–experiment relationship explicit.

Q1. How can we calculate the LL value in Theorem 6 for neural networks?

The MNIST head-update experiment uses the explicit global bound L=1/2L=1/\sqrt{2} in Frobenius norm.

For the frozen encoder, only the head matrix WW varies, and the logits are gx(W)=Wxg_x(W)=Wx. Unit-bounded augmented embeddings give B1≤1B_1\le1 and B2=B3=0B_2=B_3=0 in Frobenius parameter norm and Euclidean logit norm. Appendix C.1 derives the global negative-cross-entropy bound A3=1/2A_3=1/\sqrt{2}, so substitution permits L=1/2L=1/\sqrt{2}. Section 5.4 and Appendix C.3 apply that analytic constant to the head-update contribution game; the encoder parameters remain fixed.

For a fully updated smooth network, the regional derivative calculation supplies a sufficient bound.

First enclose the coalition models using the public checkpoint and weighted update-norm caps aia_i. At maximum common scale ε0\varepsilon_0, the triangle inequality gives a ball around the checkpoint with radius ρ=ε0∑iai\rho=\varepsilon_0\sum_i a_i. The utility must be three times continuously differentiable near that region. This construction uses public caps and requires no individual update directions or extra coalition evaluations.

Next use the public input bound, regional weight bounds and analytic activation derivatives in the chain and product rules through the layers. Let B1,B2,B3B_1,B_2,B_3 bound the network logit map's first three parameter derivatives, and let A1,A2,A3A_1,A_2,A_3 bound the utility's first three logit derivatives. The bounds must hold throughout the region and for every allowed evaluation input and label. The third-order chain rule gives

Lexample=A3B13+3A2B1B2+A1B3L_{\mathrm{example}}=A_3B_1^3+3A_2B_1B_2+A_1B_3.

A uniform example-level bound also bounds the empirical mean utility; alternatively, average separate valid example-level bounds. Pointwise automatic differentiation at the checkpoint or sampled models does not establish the regional supremum. Appendix B.4 gives this construction and discusses the conservatism of the tested full-network bounds. ReLU activation boundaries require a separate smoothness argument.

Dear Reviewer,

We have posted detailed responses to your comments and uploaded a revised manuscript. We clarified the distinction between the structural identified set and the smooth certificates used in the experiments, added a clipping and scale sensitivity analysis and an MNIST head-update study, and explained how to obtain the required derivative bound for neural networks (Appendices B.4 and C.2–C.3).

As the discussion period ends on November 18, we would be grateful to hear whether these revisions address your concerns, or whether any points would benefit from further clarification.

Thank you for your time and consideration.

We thank the authors for their detailed response.

W1: My concern has been addressed.

W2: I still do not see any multi-round results.

Q1: For a fully updated smooth network, the computed value of L appears to be too large, so large that the certificate becomes completely uninformative. I suspect this is also why the MNIST experiment updates only the linear logistic head. This raises my concerns about the practical benefit of the paper’s contributions.

Reviewer_nZUQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether Shapley values can be recovered from standalone and leave-one-out utility evaluations in federated learning. It identifies when these observations are sufficient and provides intervals and approximation guarantees when exact recovery is not possible.

Critical strengths

1.The problem is clearly defined. The paper states what information the scorer can access and gives a coherent analysis of what can be inferred from it. 2.I find the intervals useful because they can determine whether a client contributes positively and whether one client contributes more than another. This makes the results easier to use for contribution assessment than an approximation error alone.

Critical weaknesses

1.The assumption of honest reporting limits practical applicability. It may be acceptable among cooperating institutions, but becomes harder to justify when scores determine payments. The paper also assumes a common utility function without sufficiently explaining how clients would access shared evaluation data or a trusted evaluation service. 2.I am less convinced by the practical evidence for the certificates. The certification experiments use logistic regression with normalized features, clipped updates, and additional update scaling. It remains unclear whether the intervals would be informative for larger updates or more complex models, and how these restrictions affect training progress. 3.The analysis concerns a single completed round with fixed updates and aggregation weights. These scores do not directly measure an institution’s data value over the full training process. This limited scope, together with the gaps in practical validation, motivates a weak reject recommendation.

Questions for authors

pls see weeknesses

3 comments

Thank you for your time and for reviewing our paper. We have uploaded an updated version of the paper. The theorem numbering and all section, table and appendix references below refer to that version. Original Theorems 6–8 are now Theorems 5–7.

The replies below explain how the evaluator supplies the measurements, what the revised experiments establish, and why completed-round attribution is an operational task.

W1. Honest reporting is harder to justify when scores determine payments, and the common utility requires an explanation of access to shared evaluation data or a trusted evaluation service.

The trusted-evaluator arrangement in Section 2 and Appendix E removes the need for participants to declare their own utility measurements or share held-out examples. The consortium agrees on the loss, evaluation population and weights. The service holds the data, evaluates committed singleton, full and leave-one-out models internally, and supplies the authorized scalars. Under this arrangement, a client's claimed utility is not an input to its score.

The revised paper states the requirements making this arrangement consistent: bind the scored updates to those aggregated, fix the evaluator and enforce the certificate bounds. The privacy requirement also has a direct reason. Subtracting a leave-one-out model from the full model reveals that client's weighted update, so the service protects those models and restricts the released outputs. Appendix E specifies these obligations and alternative trust models. Strategic participation and payment incentives concern the mechanism built on the correctly evaluated scores.

Correct measurement does not resolve the information problem. Our construction gives two update profiles with the same endpoint measurements, checkpoint, evaluator, weights and aggregate but different Shapley vectors. The impossibility therefore persists under a trusted evaluator. Section 2 and Appendix E now separate the interface supplying correct measurements from the identifying information those measurements contain.

W2. The experiments use logistic regression with normalized features, clipping and additional scaling. It is unclear whether certificates remain informative for larger updates or more complex models, and how these restrictions affect progress.

We added a paired clipping/scale comparison and an unclipped progress reference to measure the tradeoff directly. Using the same profiles fixes the underlying updates and checkpoints while the clip and common scale vary. At full scale, the highlighted settings retain 87.6% median unclipped round gain on UCI-HAR and 86.3% on MHEALTH. Combined certificates resolve 86.1% and 84.1% of pairs, with 14/20 and 13/20 winners. At those identical updates, the coordinate rule resolves about 80% of pairs and 7/20 and 11/20 winners. The direct bound therefore adds guarantees without further reducing progress. Section 5.3, Table 1 and Appendix C.2 report the full comparison.

The MNIST study extends certification to head updates over a learned, frozen representation. Eight clients update a logistic head on unit-bounded embeddings, giving the analytic third-derivative bound. A selection rule fixed before training chooses among eight clip/scale candidates using validation endpoints and public bounds before accessing test outcomes. Validation data are disjoint from encoder training, checkpoint training, client data and test data. Section 5.4 and Appendix C.3 specify this protocol.

All 20 selections are eligible. The original coordinate rule covers every test game and certifies 526/560 pairs, 155/160 signs and 15/20 winners. Every replicate exceeds 80% pair certification: the minimum is 23/28, the median 27/28 and the maximum 28/28. The selected (clip, scale) settings are (0.1, 1) in 13 replicates, (0.1, 0.5) in six and (0.2, 0.5) in one. All unclipped reference gains are positive, ranging from 1.226×10−41.226\times10^{-4} to 1.332×10−41.332\times10^{-4}.

With those selections fixed, the combined rule certifies 533/560 pairs and 18/20 winners, with per-round minimum/median/maximum counts of 24/27/28. Median retained head-update gain is 55.1%; median utility gain is 6.95×10−56.95\times10^{-5}, and median accuracy is 88.1% at both the checkpoint and aggregate. These partition replicates share the encoder, checkpoint, image pools and evaluation sets. This study demonstrates certificate and validation-selection transfer for the head-update game; the 86–88% retention result comes from the separate logistic sensitivity.

Appendix C.7 reports the retrospective direct-pair refinement with selections unchanged. Section 5.4 and Table 5 distinguish the original coordinate results from those combined results. Appendix C.4 separately accounts for evaluation cost, including the selector's 144 validation utilities per profile.

W3. The analysis concerns one completed round with fixed updates and weights, so the scores do not directly measure an institution's data value over the full training process.

The fixed-round game evaluates the updates that produced the released model change. Its counterfactual removes a recorded update while keeping the checkpoint and aggregation weights fixed. This supplies a post-round audit with concrete decisions: under the stated controls, an interval above zero certifies a positive Shapley contribution, and a positive lower bound on a pair difference certifies the ordering. The information results determine when the authorized measurements support those conclusions.

Removing an institution throughout training changes later checkpoints and the other clients' updates. Summing the completed-round scores therefore generally differs from full-retraining institutional data Shapley. Section 2 and the conclusion now state this distinction explicitly and separate the fixed-update game from retraining and subset-dependent weight normalization.

Dear Reviewer,

We have posted detailed responses to your comments and uploaded a revised manuscript. We clarified the trusted-evaluator arrangement and the scope of completed-round attribution, added a sensitivity analysis quantifying certification and retained learning progress, and extended the certificate experiments to logistic-head updates over a learned MNIST representation (Appendices C.2–C.3).

As the discussion period ends on November 18, we would be grateful to hear whether these revisions address your concerns, or whether any points would benefit from further clarification.

Thank you for your time and consideration.

Reviewer_Qsy2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies whether Shapley values for a completed, fixed-update federated learning round can be recovered from the grand-coalition worth together with each client’s standalone gain and leave-one-out marginal contribution. It shows that the midpoint of the two client-specific measurements is exact for 2-additive games. For four or more clients, however, positive 3-additive games can have identical observations and different Shapley values, even when the public evaluator and released aggregate are fixed.

The paper then characterizes the complete set of compatible Shapley vectors under positive third-order interactions, including sharp coordinate intervals and robust rankings. For smooth utilities, it derives an explicit cubic error bound for the midpoint and a matching worst-case lower bound on the attainable error order. Experiments examine the predicted local scaling on Fashion-MNIST and demonstrate decision certificates in controlled eight-client logistic-regression settings on UCI-HAR and MHEALTH.

Critical strengths
  1. The paper makes a useful distinction between computational difficulty and lack of identifying information. Its impossibility result is stronger than a dimension-counting argument: the construction preserves the endpoint observations, public evaluator, checkpoint, weights, and aggregate while changing the Shapley vector. This establishes an actual limitation of the permitted information. Theorem 6 complements this negative result with an exact characterization of the compatible Shapley vectors under positive 3-additivity, including attainable bounds and necessary-and-sufficient ranking conditions. I consider this combination a substantive theoretical contribution.

  2. The smooth analysis provides both an explicit finite-scale guarantee and a matching worst-case error order. Theorem 7 identifies the mixed third-order interactions responsible for the midpoint’s error, while Theorem 8 shows that the cubic order cannot generally be improved using the same information. The empirical evaluation also clearly separates the endpoint information used by the scorer from the full-coalition evaluations used to audit its outputs. Fixing the certificate protocol before the MHEALTH evaluation, and reporting the unfavorable structural-assumption check in Appendix C.3, make the scope of the evidence transparent.

Critical weaknesses
  1. The practical relevance of the sharp structural identified set remains insufficiently established. Theorem 6 requires nonnegative Harsanyi dividends and exact third-order additivity. These conditions are substantially stronger than positive overall utility improvement. Appendix C.3 reports that only 6 of 273 evaluated endpoint observations admit a positive 3-additive extension, and that the exact Shapley vector of each of those six underlying games falls outside at least one resulting structural interval. I appreciate that the authors report this finding and correctly distinguish compatibility of an observation from membership of the true game in the assumed class. Nevertheless, the experiments do not demonstrate an FL setting in which these structural intervals can safely be applied to the original game. This limits the operational significance of a central result unless the assumptions can be independently justified using information available to the protocol. My concern is about applicability, not the validity of the conditional characterization.

  2. The very high decision-certification rates are demonstrated only in a tightly controlled update regime, and the cost of that control is not quantified. The UCI-HAR and MHEALTH experiments use logistic models with normalized features, an analytic third-derivative bound, local-update clipping at 0.2, and a further common scale of 0.5. Appendix C.1 states that all 160 MHEALTH updates reach the clipping threshold. This is a valid demonstration of the smooth certificate. However, its radius scales cubically with the update scale, while the Shapley signal is first order under the theorem’s nonzero-signal condition. Consequently, reducing the update scale improves relative certifiability by construction. The present results do not establish how much learning progress is retained when obtaining the reported certification rates, or how informative the intervals remain under less restrictive updates. The Fashion-MNIST experiment verifies the local error scaling but does not provide a corresponding neural-model certificate. A focused sensitivity analysis within the existing logistic setup would substantially strengthen the practical claim.

Questions for authors
  1. Clarification concerning Theorem 6: Can you identify a concrete FL setting in which nonnegative dividends and third-order additivity can be established independently of the unobserved coalition values? Given the results in Appendix C.3, what would justify applying the structural identified set to an actual round beyond checking that its endpoint observation admits some positive 3-additive extension?

  2. Evidence concerning the existing certificate experiment: At ε = 1, and at a modestly less restrictive clipping threshold, how do the certified-pair fraction, interval widths, and grand-coalition utility gain change relative to the protocol in Table 1? A small sensitivity analysis using the current setup would clarify whether the high certification rates are compatible with meaningful learning progress. This request concerns the operating range of the existing result, rather than a new large-scale benchmark.

3 comments

Thank you for taking the time to review our paper. We have uploaded an updated version of the paper. The theorem numbering and all section, table and appendix references below refer to that version. Original Theorems 6–8 are now Theorems 5–7.

The replies below separate the structural information limit from the operational certificates and quantify the progress retained by those certificates.

W1. The structural intervals require nonnegative Harsanyi dividends and exact third-order additivity, and the experiments do not independently justify applying them to the true FL game.

Identical endpoints can conceal different top contributors even with nonnegative effects. Figure 1 gives two four-client games with si=1s_i=1, mi=7m_i=7 and V=14V=14, but Shapley vectors (4,10/3,10/3,10/3)(4,10/3,10/3,10/3) and (10/3,10/3,10/3,4)(10/3,10/3,10/3,4). Moving effects between pairs and triples preserves the measurements while changing who receives the largest share.

Theorem 5 resolves the resulting uncertainty under positive third-order structure: it gives every attainable vector, sharp coordinate bounds and necessary-and-sufficient robust rankings. This is the first interaction class beyond pairwise exact recovery. The experimental certificates instead bound attribution error for signed interactions of arbitrary order using public curvature and update controls.

Endpoint feasibility establishes compatibility; applying the structural intervals requires independent membership justification. Appendix C.6 reports 6/273 compatible observations, each with a true Shapley vector outside at least one interval. Section 3.3 identifies the set as a structural benchmark, and Section 4.1 supplies the operational certificate used in the experiments.

W2. The original high certification rates use normalized logistic models, clipping and additional update scaling. Their progress cost is unquantified, and the Fashion-MNIST scaling study does not provide a neural-model certificate.

At full scale, UCI-HAR with clip 0.2 retains 87.6% median unclipped round gain and certifies 482/560 pairs with the combined rule; MHEALTH with clip 0.4 retains 86.3% and certifies 471/560. At unchanged updates, certified winners increase from 7/20 to 14/20 and from 11/20 to 13/20, respectively.

The added comparison uses the same 20 profiles per dataset at clips {0.2,0.4}\{0.2,0.4\} and scales {0.5,1}\{0.5,1\}. Fixing checkpoints, training, feature normalization and weights isolates the effect of those controls. Each setting uses its public cap; all paired unclipped reference gains are positive.

DatasetClipScaleMedian round gain retainedCoordinate widthCoordinate pairsCombined pairsWinners /20 (coordinate → combined)
UCI-HAR0.20.544.1%2.62%539/560545/56018 → 18
UCI-HAR0.2187.6%10.54%446/560482/5607 → 14
UCI-HAR0.40.550.4%18.46%294/560362/5601 → 4
UCI-HAR0.41100.0%74.52%3/56031/5600 → 0
MHEALTH0.20.522.6%1.24%555/560555/56020 → 20
MHEALTH0.2145.1%4.98%534/560542/56017 → 18
MHEALTH0.40.543.5%5.22%528/560542/56015 → 17
MHEALTH0.4186.3%21.01%448/560471/56011 → 13

Retention is the median within-profile ratio to the unclipped gain; width is median total efficiency-tightened coordinate width divided by the Shapley ℓ1\ell_1 norm. Full-scale median gains are 0.005847 and 0.023017. Section 5.3, Table 1 and Figure 2 summarize the comparison; Appendix C.2 gives the full sensitivity.

Interactions shared by two clients contribute equally to their midpoint errors and cancel in the comparison. Corollary 16 uses this cancellation; intersecting its pair interval with the efficiency-tightened coordinate difference retains all earlier decisions without extra utility queries. This refinement is retrospective at fixed settings; original confirmation and selection rules are unchanged. Figure 2 shows the original coordinate-only rule.

Raw midpoint rankings and winners agree with exact Shapley in all 200 rescored logistic and neural-head game evaluations. The combined bound guarantees more of those correct rankings from the same information. A winner requires certification against every competitor; other pairs remain unresolved. Appendix C.7 reports the audit.

On MNIST, eight clients update a logistic head over a learned, fixed encoder. At unchanged validation-selected settings, the combined rule raises test certification from 526/560 to 533/560 pairs and from 15/20 to 18/20 winners, with 55.1% median retained head-update gain. Section 5.4 and Appendix C.3 report this head-only study. The abstract and contribution list now highlight 84–86% combined pair certification at 86–88% median gain retention from the separate logistic operating points.

Q1. Can you identify a concrete FL setting in which nonnegative dividends and third-order additivity can be established independently of the unobserved coalition values? Given the results in Appendix C.6, what would justify applying the structural identified set beyond checking endpoint compatibility?

An actual-round application requires independent justification of nonnegative dividends and zero dividends above order three; endpoint compatibility alone does not supply it. We do not establish a natural-FL membership guarantee. Theorem 5 gives the conditional characterization under W1; the demonstrated smooth certificates require no positive third-order membership. Sections 3.3 and 4.1 now separate these roles.

Q2. At ε=1\varepsilon=1 and a modestly less restrictive clipping threshold, how do the certified-pair fraction, interval widths and grand-coalition utility gain change relative to the original protocol?

The full-scale UCI-HAR setting at clip 0.2 has median gain 0.005847, 87.6% retained gain and 10.54% relative coordinate width; it certifies 446/560 pairs with the original rule and 482/560 with the combined rule. MHEALTH at clip 0.4 has median gain 0.023017, 86.3% retention and 21.01% width; the corresponding pair counts are 448/560 and 471/560. W2 gives the original half-scale settings and the full grid, so the progress and certification changes are directly comparable within the same profiles.

Dear Reviewer,

We have posted detailed responses to your comments and uploaded a revised manuscript. We clarified the conditional role of the structural identified set and added the requested clipping and scale sensitivity analysis (Appendix C.2). At full scale, the highlighted logistic settings certify 84–86% of client pairs while retaining 86–88% of the median unclipped round gain.

As the discussion period ends on November 18, we would be grateful to hear whether these revisions address your concerns, or whether any points would benefit from further clarification.

Thank you for your time and consideration.

Comments on the paper

1 comment

We thank all the reviewers for assessing our work. To improve the work, we have uploaded a revised manuscript addressing the reviews:

  • Added a paired sensitivity analysis showing 84–86% certified pairwise rankings at settings retaining 86–88% median within-profile round gain relative to unclipped updates.
  • Derived a direct pairwise certificate that improves guaranteed rankings and winner identification without additional utility evaluations.
  • Added an audit distinguishing observed midpoint ranking accuracy from certified decisions.
  • Added a MNIST experiment demonstrating certification for head updates over a frozen learned representation, with validation-only setting selection.
  • Expanded the evaluation-interface and trust-model explanation, clarified completed-round attribution and the structural theorem’s applicability, and provided the neural derivative-bound construction.
  • Corrected the three-client edge case and improved definitions and cross-references.

We hope these additions and clarifications address the reviewers’ concerns. The individual responses below explain how each comment has been addressed and identify the corresponding changes in the revised paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.