acceptodds
Under review as a conference paper at ICLR 2027

FLARE: Reliable Local Explanations through Federated Representation Learning

Abstract

Federated learning enables collaborative model training without sharing raw data, but local explanation methods remain constrained by the data available at individual clients. When local data are sparse, imbalanced, or heterogeneous, query neighborhoods may provide insufficient support for representative and faithful explanations. Although federated explainability methods can exploit distributed information in different ways, the use of collaboratively learned representations to improve real-instance, query-specific neighborhood construction for local surrogate explanations remains largely unexplored. We propose FLARE, a federated framework for reliable local explanations. To the best of our knowledge, FLARE is the first framework to combine federated representation learning with client-local, real-instance neighborhoods for query-specific surrogate explanations. A collaboratively trained SCARF encoder guides local source–target neighborhood construction, while support-aware neighborhood completion and adaptive surrogate weighting improve reliability. Explanations are generated using weighted Ridge surrogates and evaluated for fidelity, support, stability, and robustness. We evaluate FLARE across five tabular datasets spanning seven classification tasks, including predefined source–target domain shifts and natural multi-client partitions, under varying client counts, federated optimizers, and black-box predictors. FLARE achieves the strongest overall fidelity across the evaluated baselines, with sensitivity analyses showing that it remains effective across different client counts and black-box predictors. These results show that cross-client structural knowledge can improve local explanation construction while keeping raw data and explanation generation client-local.

Primary area:
interpretability and explainable AI
Keywords:
Federated Learning, Explainable AI, Privacy-preserving, Local Explanations, Federated Explainability

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_K5m7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes FLARE, a federated framework for generating reliable local explanations using collaboratively learned SCARF representations while keeping raw data and explanation generation client-local. It constructs query-specific neighborhoods from real source and target instances, incorporates support-aware neighborhood completion and adaptive weighting, and fits weighted Ridge surrogates. Experiments across seven classification tasks compare FLARE against federated explainability baselines, reporting improvements in fidelity, reliability, and robustness across different client configurations and black-box models.

Critical strengths
  1. The paper addresses an important problem in federated explainability, particularly the difficulty of generating faithful local explanations when client data are sparse or heterogeneous.
  2. The evaluation considers multiple datasets, client configurations, and black-box models, incorporating reliability measures beyond conventional fidelity, including neighborhood support, stability, and robustness.
Critical weaknesses
  1. The contribution of federated representation learning is not adequately isolated. Comparisons against locally trained SCARF or non-federated representation baselines are missing, making it unclear whether federation itself drives the improvements.
  2. Baseline comparability is questionable because FLARE uses client-local black-box predictors, while some baselines explain federated predictors. The ablation study also disables two proposed mechanisms simultaneously, preventing attribution of their individual contributions.
  3. Several reporting inconsistencies undermine confidence in the results, including contradictory FedProx coefficients, identical FastSHAP++ results across client counts, and claims of superiority on every metric that are not fully supported by Table 1.
Questions for authors
  1. Can the authors compare federated SCARF with locally trained SCARF and conventional feature-space representations under otherwise identical settings?
  2. Are all methods evaluated against identical black-box predictors and common evaluation neighborhoods? If not, how are the fidelity comparisons justified?
  3. Can the authors provide separate ablations for neighborhood completion and adaptive weighting, along with baseline comparisons using the same reliability metrics?
  4. Can the authors clarify the inconsistent FedProx coefficients, repeated FastSHAP++ results across client counts, and discrepancies between the claimed and reported performance?
5 comments

We thank the reviewer for the feedback and detailed observations. We address the concerns below and provide additional experimental details.

1. Contribution of federated representation learning

We conduct additional experiments with independently trained client-local SCARF encoders (LocalEnc). We compare LocalEnc with FedAvg (FA) and FedProx (FP) across 3, 5, and 10 clients using SVM and DNN black-box models.

The following table presents representative results for the 5-client setting using SVM.

DatasetLocalEnc F1/MCC/AUCFA F1/MCC/AUCFP F1/MCC/AUC
Adult.818/.749/.974.806/.728/.967.874/.813/.987
Bank.677/.677/.976.661/.654/.973.889/.890/.998
STD1.956/.891/.992.954/.889/.992.955/.887/.992
STD2.960/.891/.993.966/.899/.994.968/.905/.993
DTD1.343/.339/.895.252/.251/.835.818/.821/.997
DTD2.399/.395/.864.273/.268/.796.891/.887/.997
ACS (50).846/.853/.880.745/.650/.910.850/.860/.960

All datasets use five clients except ACS Income (50 clients).

The results show that federated encoder training, particularly with FedProx, generally provides stronger fidelity than local training, with more improvements on Adult, Bank and Diabetes datasets. LocalEnc remains competitive on Student datasets and ACS Income. These findings support the benefits of collaborative representation learning without sharing raw client data.

2. Component-wise ablation study

We agree that disabling both mechanisms simultaneously does not isolate their individual contributions. Therefore, we conduct additional experiments evaluating adaptive surrogate class weighting (CW) and support-aware neighborhood completion (NC) separately and jointly.

The following table presents Fidelity F1 and explanation reliability (Rel.) for the 5-client FedProx-SVM setting.

DatasetNeither F1/Rel.CW only F1/Rel.NC only F1/Rel.Both F1/Rel.
Adult.742/.450.829/.470.793/.660.874/.740
Bank.674/.430.738/.430.774/.630.889/.650
STD1.935/.750.954/.760.932/.880.955/.980
STD2.957/.790.968/.800.953/.920.968/.980
DTD1.189/.080.415/.070.510/.330.818/.420
DTD2.275/.120.478/.100.630/.460.891/.550

The complete configuration achieves the highest Fidelity F1 and the highest explanation reliability across all datasets. CW generally improves fidelity, whereas NC more consistently improves reliability. Their combination provides the strongest overall results. We will incorporate these separate ablations into the revised manuscript.

3. Baseline comparability and evaluation protocol

FLARE is evaluated using SVM and DNN predictors, while FedRF-Surrogate, Fed-LIME, and FastSHAP++ results in Table 1 correspond to SVM black-box configurations. We acknowledge that the current presentation does not sufficiently establish whether the predictors and evaluation neighborhoods are identical across methods. We are further working on analyzing the black-box predictor configurations and evaluation protocols and will provide additional clarifications and results in a follow-up response.

We thank the reviewer again for the valuable suggestions. The additional experiments strengthen the evaluation of federated representation learning and the individual components of FLARE. We are also working on the remaining concerns and would be happy to provide further results or clarifications to support your final assessment.

AuthorsOpenReview ↗

Final days

Dear Reviewer, as the reviewer discussion phase comes to an end tomorrow, we hope that we have adequately addressed your comments and concerns. Please let us know if you need any further clarification or information to help you reassess your evaluation. Thank you again for your time and valuable feedback.

AuthorsOpenReview ↗

Consistency of Black-Box Predictors and Common Evaluation Neighborhoods

To address this concern, we ensure that all baseline methods use the same G-LibSVM black-box model configuration as FLARE. We clarify that the original baseline implementations employ different black-box prediction models. During manuscript preparation, we extend these implementations to support G-LibSVM that allows all methods to explain the same client-specific prediction targets while preserving their respective explanation mechanisms. Furthermore, all methods explain identical target-test instances and are evaluated on common held-out neighborhoods. We compute fidelity F1, MCC and AUC using a consistent evaluation protocol.

To verify this setup, we conduct prediction consistency checks across multiple test instances for each client. Specifically, we provide identical inputs to the G-LibSVM predictors used by FLARE and the baseline implementations and compare their predicted class labels and class probabilities. We confirm agreement across the evaluated instances. We also verify that all methods explain the same queries and use identical held-out evaluation observations by matching client identifiers, query indices and evaluation sample indices. These checks confirm that all methods explain consistent black-box predictions and are evaluated on the same test instances and neighborhoods, thereby allowing a fair comparison of their explanation fidelity.

We'll further clarify this evaluation protocol in the revised manuscript.

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

FLARE is a federated framework for local surrogate explanations on tabular data. Clients train a shared SCARF encoder with FedAvg or FedProx, keeping Batch Normalization local, and each query neighborhood combines Euclidean target-training neighbors with source instances from a class-compatible K-medoids cluster. Latent-locality weights, support-aware neighborhood completion, and adaptive class-support weighting feed a weighted Ridge surrogate. Experiments cover seven binary tasks from five datasets, including domain shifts and a 50-state ACS Income partition, with 3, 5, and 10 clients, SVM and DNN black boxes, and three federated baselines.

Critical strengths

S1: The paper addresses a practical gap. FLARE builds neighborhoods from real client instances instead of perturbations, keeps black-box queries and surrogate fitting local, and shares only SCARF parameters with local Batch Normalization, a clean design that Algorithm 1 and Figure 1 make easy to follow.

S2: The reliability protocol is thoughtful. Held-out fidelity, support rate, JC, and LLE are defined in Section 3.4, and Eq. (8) reweights only the fit while fidelity keeps v_i. Table 2 shows why this matters: Bank DNN reaches Support 1.000 but Reliable only 0.070.

S3: The experiments are broad and the components pay off. Seven tasks, a 50-state ACS Income federation, 3, 5, and 10 clients, and two black boxes are covered, and Table 3 shows Diabetes-DTD1 F1 dropping from 0.818 to 0.189 without the reliability components.

Critical weaknesses

W1: Section 4.4 averages five runs, but Tables 1 to 3 give no standard deviations and Table 5 lists a single seed 42. Some gaps are small, such as Student-STD2 F1 of 0.968 for FP-SVM against 0.966 for FA-SVM, so per-run spread or confidence intervals would make the rankings easier to trust.

W2: The reliability settings are fixed at one value each, τs = 0.10, q = 0.95, γ = 1.25, and ρ = 0.15, and the kernel width σ in Eq. (5) and the Ridge λ in Eq. (9) are not reported. A sensitivity sweep on the Diabetes tasks, where Table 3 shows completion matters most, would show how robust the gains are.

W3: No runtime or communication cost is reported. FLARE trains SCARF for 20 rounds with five local epochs and then runs neighborhood construction, completion, and Ridge fitting per query, plus repeated explanations for JC and LLE. Reporting encoder size, bytes per round, and per-query latency would clarify deployability in the 50-client ACS Income setting.

Questions for authors

Q1: Could the authors add a client-local SCARF encoder or an input-space kernel variant, since Section 3.2 retrieves neighbors in input space? It would also help to explain why FP-SVM far exceeds FA-SVM on Diabetes-DTD1 with μ = 0.01, while Table 4 lists 0.1.

Q2: Are all methods scored on the same points? In Table 1 FastSHAP++ has higher Student-STD1 MCC, 0.916 against 0.889, and the best baseline F1 on Diabetes-DTD2 is 0.876 rather than 0.396, so Section 5.1 could be reworded.

Q3: Minor: Table 7 is missing; Fed-LIME and FL-LIME are both used; Table 8 and Table 9 captions say 5 clients; "andρ" appears on line 269.

Q4: Related Work could also cover self-explaining rationalization, which grounds explanations in selected input evidence instead of a post-hoc surrogate and so offers a complementary route to the faithfulness FLARE targets, for example [1][2].

[1] Is the MMI Criterion Necessary for Interpretability? Degenerating Non-causal Features to Plain Noise for Self-Rationalization.

[2] Breaking Free from MMI: A New Frontier in Rationalization by Probing Input Utilization.

3 comments

We thank the reviewer for the positive assessment and suggestions. We address the questions below.

Q1. Local SCARF encoders and FedProx coefficient

We conduct additional experiments with independently trained client-local SCARF encoders (LocalEnc), without federated encoder aggregation. We compare LocalEnc with FedAvg (FA) and FedProx (FP).

The following table presents representative 5-client SVM results.

DatasetLocalEnc F1/MCC/AUCFA F1/MCC/AUCFP F1/MCC/AUC
Adult.818/.749/.974.806/.728/.967.874/.813/.987
Bank.677/.677/.976.661/.654/.973.889/.890/.998
STD1.956/.891/.992.954/.889/.992.955/.887/.992
STD2.960/.891/.993.966/.899/.994.968/.905/.993
DTD1.343/.339/.895.252/.251/.835.818/.821/.997
DTD2.399/.395/.864.273/.268/.796.891/.887/.997
ACS (50).846/.853/.880.745/.650/.910.850/.860/.960

All datasets use five clients except ACS Income (50 clients).

The results show that federated encoder training, particularly with FedProx, generally improves fidelity over local encoder training, with substantial gains on the Diabetes datasets. LocalEnc remains competitive on the Student datasets and ACS Income. These findings support collaborative representation learning without sharing raw client data.

We will verify and clarify the discrepancy between the FedProx coefficients reported in the methodology and Table 4, including the precise value used for the Diabetes-DTD1 experiments.

Q2. Evaluation comparability and reported performance

We also agree that the performance statements in Section 5.1 require correction. FastSHAP++ achieves a higher MCC on Student–STD1 (0.916 versus 0.889 for FLARE FA-SVM), and the strongest baseline F1 on Diabetes–DTD2 is 0.876. We will revise the discussion to accurately reflect these comparisons rather than claiming superiority across every metric.

Q3. Presentation and consistency

We will correct the missing Table 7 reference, consistently use Fed-LIME throughout the manuscript, revise the captions of Tables 8 and 9 to reflect the evaluated client configurations and correct other the formatting errors.

Q4. Related work on self-explaining rationalization

We agree that self-explaining rationalization provides a complementary perspective on explanation faithfulness by identifying informative input evidence rather than relying on a post-hoc surrogate. We will add more papers in Related Work to discuss the proposed approaches related to MMI Criterion, and clarify how their rationale-selection objectives differ from FLARE's local surrogate-based explanation framework.

We thank the reviewer again for these suggestions, which help us improve the empirical evaluation, methodological clarity, and presentation of FLARE. We would be happy to provide further clarifications or additional results if needed.

Thank you for the authors' response. I will maintain my score.

Dear Reviewer, is there anything we can execute and/or run to make you change from weak acceptance to clear acceptance? We would be happy to improve our research

Reviewer_NTwkOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FLARE, a federated framework that leverages collaborative representation learning to improve local explanations. An encoder is trained via FedAvg/FedProx, and used to weight the neighborhood of each client's local data prior to fitting a weighted ridge surrogate to the black box predictions. Class imbalances within each neighborhood are addressed by support-aware completion and adaptive minority-class reweighting. The method is evaluated on a series of tabular tasks against other federated explainability baselines.

Critical strengths
  • The problem of federated XAI is compelling, and work leveraging federated knowledge to improve explanations deserves attention. This paper presents promising results and public benchmark code which supports future work in the area.
Critical weaknesses
  • The federated component of this work is underdeveloped. The benefits of federation come from the SCARF encoder, but it is unclear how the FedAvg/FedProx models compare to locally trained encoders, or to an ideal centralized encoder. Applying existing FL techniques to a new problem without further analysis limits the impact of the work.
  • The flow of the paper is difficult to follow, with many details left unclear and notation often inconsistent. The method appears largely heuristic, and explanations/evidence to support them are not provided (e.g., why use two different distance spaces, choice of surrogate, …). The evaluation setup has similar issues (e.g., what black box model are the baselines explaining).
Questions for authors
  • Could you evaluate FLARE using local-only SCARF encoders? This could help highlight the benefits of federation and how aspects of FL scenarios impact the learned explainer (e.g., the effect of client heterogeneity, choice of FL algorithm)
  • Could you clarify which black box model each method explains in Table 1?
4 comments

We thank the reviewer for the constructive feedback. We address the concerns related to federated representation learning and experimental clarity below.

1. Contribution of federated representation learning

Following the reviewer's suggestion, we conduct additional experiments with locally trained SCARF encoders (LocalEnc). We compare LocalEnc against FLARE with FedAvg (FA) and FedProx (FP) across 3, 5, and 10 clients using SVM and DNN black-box models.

The following table presents representative 5-client results for SVM.

Additional evaluation: Federated vs. locally trained SCARF encoders

We evaluate FLARE with independently trained local SCARF encoders (LocalEnc), without federated encoder aggregation. We compare LocalEnc with FedAvg (FA) and FedProx (FP) using the same SVM black-box model.

DatasetLocalEnc F1/MCC/AUCFA F1/MCC/AUCFP F1/MCC/AUC
Adult.818/.749/.974.806/.728/.967.874/.813/.987
Bank.677/.677/.976.661/.654/.973.889/.890/.998
STD1.956/.891/.992.954/.889/.992.955/.887/.992
STD2.960/.891/.993.966/.899/.994.968/.905/.993
DTD1.343/.339/.895.252/.251/.835.818/.821/.997
DTD2.399/.395/.864.273/.268/.796.891/.887/.997
ACS (50).846/.853/.880.745/.650/.910.850/.860/.960

Values represent Fidelity F1/MCC/AUC. All datasets use five clients except ACS Income (50 clients).

The results demonstrate that federated encoder training, particularly with FedProx achieves competitive results. On Student–STD1, LocalEnc achieves marginally higher F1 and MCC, whereas FedProx achieves higher F1 and MCC on Student–STD2. The improvements from federation are better visible on Adult, Bank and especially the Diabetes datasets. These findings highlight the benefits of federated representation learning, which enables collaborative training without sharing raw client data while maintaining competitive explanation fidelity and reliability and often outperforming locally trained encoders.

2. Methodological motivation and ablation

To further validate the proposed components, we conduct a four-way ablation study of adaptive surrogate class weighting (CW) and support-aware neighborhood completion (NC): neither component, CW only, NC only and both components. The following table reports Fidelity F1 and other metrics in the 5-client FedProx-SVM setting.

DatasetNeither F1/Rel.CW only F1/Rel.NC only F1/Rel.Both F1/Rel.
Adult.742/.450.829/.470.793/.660.874/.740
Bank.674/.430.738/.430.774/.630.889/.650
STD1.935/.750.954/.760.932/.880.955/.980
STD2.957/.790.968/.800.953/.920.968/.980
DTD1.189/.080.415/.070.510/.330.818/.420
DTD2.275/.120.478/.100.630/.460.891/.550

The complete configuration achieves the better Fidelity F1 and the highest explanation reliability across all six datasets. The individual components also provide benefits: CW generally improves fidelity, whereas NC more consistently improves reliability. Their combination shows the strongest overall results. We will further clarify the motivation for the two distance spaces, surrogate selection, and mathematical notation in the revised manuscript.

3. Black-box models and evaluation clarity

FLARE explains SVM and DNN predictive models, as indicated by the corresponding configurations. FA and FP refer to the federated encoder aggregation algorithms. The other baselines also use SVM black-box models. We will clarify these configurations in the revised manuscript.

We thank the reviewer again for the constructive suggestions, which help us strengthen the empirical evaluation and presentation of FLARE. We would be happy to provide additional results or clarifications if needed to further address your concerns and support your final assessment.

AuthorsOpenReview ↗

Final days

Dear Reviewer, as the reviewer discussion phase comes to an end tomorrow, we hope that we have adequately addressed your comments and concerns. Please let us know if you need any further clarification or information to help you reassess your evaluation. Thank you again for your time and valuable feedback.

Dear Reviewer NTwk,

As the discussion is nearing its end, we kindly invite you to take a fresh look at the revised manuscript and our responses. If, upon reconsideration, you feel that the paper has improved in line with your expectations and that your earlier concerns have been adequately addressed, we would be thankful if you could consider adjusting your score accordingly.

We sincerely thank you for your time and thoughtful review.

Best regards

Thank you for your response, the new experiments are very helpful.

I still have some mild concerns about the presentation of the final work, but these results addressed my primary concerns.

I will raise my score to a 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.