acceptodds
Withdrawn from ICLR 2027

Latent Entity Negotiation for Federated Learning with Private Taxonomies

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated classifiers normally assume that output coordinate denotes the same class at every client. This interface breaks when institutions retain taxonomy-local labels that independently permute, merge, or conditionally split latent entities; here “private” means unshared identities, not differential privacy. We introduce FedLENS, an architecture that factors each private classifier into a shared latent-entity bank, a client-private entity-to-word channel, and an input-dependent residual reader. Clients communicate adaptive micro-prototypes without word identities, and the server negotiates anonymous latent entities through mass-aware semi-unbalanced transport with learned capacities. The residual preserves feature-conditioned refinements, while the entity path transfers structure through the encoder even though the default test-time reader is private. We prove vocabulary-permutation invariance, conditional local identifiability under separated-support, pure-component, and correct-capacity assumptions, and conditional encoder convergence under stochastic-gradient local updates with explicit private-state and bank-value variation terms. Across four controlled benchmarks, negotiation improves five-seed mean accuracy by 5.00–7.62 points over the matched residual-only control; gains over the strong native-compatible FedRoD comparator are smaller, at 0.23–1.05 points.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
federated learning, representation learning, heterogeneous label spaces, private taxonomies, personalized federated learning, optimal transport

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Yi3eOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies federated learning with incompatible client label taxonomies, where labels may be permuted, merged, or split across clients. It proposes FedLENS, which learns a shared bank of anonymous latent entities from client micro-prototypes via semi-unbalanced optimal transport, while retaining client-specific mappings and residual classifiers. Experiments on four controlled benchmarks show that FedLENS improves over residual-only training and several existing baselines, with ablations evaluating the contributions of its main components.

Critical strengths

The paper addresses an important and relatively underexplored FL setting. Most federated learning methods implicitly assume that classifier coordinates have consistent semantic meanings across clients. FedLENS explicitly considers more challenging taxonomy mismatch, including label permutation, merging, and input-dependent splitting, which is a meaningful problem formulation for heterogeneous institutions and decentralized learning settings.

The combination of anonymous micro-prototypes, private label-to-entity mappings, adaptive latent capacities, and semi-unbalanced OT provides a concrete mechanism for cross-client coordination without explicitly sharing label identities. The paper also includes a relatively comprehensive set of ablations and evaluates the method across multiple controlled benchmarks, which helps demonstrate that the proposed framework can operate under the considered taxonomy mismatches.

Critical weaknesses

The necessity of the latent-entity pathway is not sufficiently established. Both the entity and residual pathways operate on the same shared representation, yet removing the entity path causes only a marginal performance drop, while removing the residual path leads to a substantial degradation. Since the residual pathway is also the default prediction path at deployment, it remains unclear what unique benefit the latent-entity negotiation provides beyond an auxiliary training signal.

The method relies on a strong assumption that semantic entities are recoverable from feature similarity. The entity bank is learned without semantic supervision, and prototype--entity matching is primarily based on feature similarity. Under substantial domain shift or heterogeneous taxonomies, feature similarity may not reliably indicate semantic identity. Thus, it is unclear whether the learned entities correspond to meaningful shared semantics or merely arbitrary clusters in the representation space.

The practical benefit of the additional optimization machinery is not sufficiently justified. FedLENS introduces prototype clustering, semi-unbalanced OT, iterative entity-bank updates, and multiple auxiliary losses and hyperparameters, while the improvements over strong baselines are relatively modest. The paper does not report training-time, convergence, or computational-cost comparisons, making it difficult to assess whether the additional complexity is warranted.

Questions for authors

What unique role does the latent-entity pathway play beyond the client-specific residual pathway? Since both pathways operate on the same shared representation and the default deployment uses the residual pathway ρm=1\rho_m=1, can the authors provide evidence that the latent-entity negotiation produces a qualitatively different or otherwise necessary form of cross-client representation transfer, rather than acting primarily as a weak auxiliary training signal?

How robust is the latent-entity negotiation to imperfect feature alignment and early-training matching errors? Since entity matching is based on prototype--entity feature similarity without semantic supervision, incorrect matches could potentially be propagated through subsequent rounds. Can the authors provide experiments or analysis showing robustness to noisy prototypes, different bank initializations, or deliberately perturbed early-round representations?

What is the empirical basis for the capacity rule Kb=max⁡(8,2M)K_b=\max(8,2M)? Since the number of clients does not necessarily determine the number of shared latent entities, can the authors clarify the rationale for this scaling and report performance under substantially under- and over-provisioned capacities?

Reviewer_rVALOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The proposed algorithm called FedLENS, aims to share useful representations without requiring a common label vocabulary and propose an algorithm to achieve it, which looks interesting, but the current manuscript is difficult to follow due to inconsistent notation, missing definitions, and insufficient explanation of the proposed method.

Critical strengths

This work proposes a federated learning algorithm that shares useful representations across clients with different private label systems.

Critical weaknesses

While the idea is interesting, the manuscript lacks the clarity needed to understand the algorithm’s design and underlying rationale.

Questions for authors
  1. The manuscript is difficult to follow due to unclear notation, missing definitions, and insufficient explanation. Readers are required to infer too many connections between the probabilistic formulation and its implementation. For example, Section 4.2 denotes the encoder function by fθf_\theta, whereas Section 4.5 refers to the broadcast encoder through its parameters θt\theta^t. Moreover, pˉmy\bar p_{my} is introduced in Section 4.5 without a precise definition. Although the channel Am(y∣z,x)A_m(y\mid z,x) appears in the introduction and the probabilistic formulation in Section 3, its relationship to the implemented decoder remains unclear.

  2. BmB_m appears to parameterize a standalone private classifier rather than an explicit residual correction to the entity predictor. The authors should clarify the term “residual” or replace it with a more descriptive term, such as “private classifier.”

  3. The motivation for using the logit-adjusted prediction q~mres(y∣x)\widetilde q_m^{\mathrm{res}}(y\mid x) during training should be explained. In particular, the authors should discuss why the adjustment is included during training, and how this choice relates to local class imbalance and the reported evaluation metrics.

  4. A table summarizing the purpose of each term in Eq. (8), the parameters receiving gradients from it, and the quantities held fixed would substantially improve clarity.

  5. Table 2 shows modest gains over FedRoD, particularly on the more complex benchmarks. Comparisons of total communication, training time, and memory usage against FedRoD would help assess whether these improvements justify the additional complexity.

Reviewer_9NHFOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes FedLENS, a federated learning framework that enables collaborative representation learning across clients with heterogeneous and private label spaces. The framework considers challenging label heterogeneity scenarios, including label permutations, merges, and input-dependent splits. FedLENS employs a shared encoder and latent entity bank, together with private entity-to-label channels and residual readers, to facilitate knowledge sharing without requiring a common label taxonomy. It further introduces adaptive micro-prototypes and mass-aware semi-unbalanced optimal transport for latent entity alignment. The paper also presents theoretical analyses and experimental evaluations across four benchmarks.

Critical strengths
  1. The paper investigates a meaningful and practically relevant problem in federated learning. In real-world collaborative learning scenarios, different organizations may maintain distinct label taxonomies, and sharing label semantics may be undesirable due to privacy or confidentiality concerns.

  2. The paper considers several challenging forms of label heterogeneity, including label permutations, class merges, and input-dependent splits.

  3. The experimental evaluation includes different comparison mechanisms, ablation studies, and stress tests. These experiments provide useful evidence for examining the contributions of different architectural components and understanding the behavior of the proposed method under different forms of label heterogeneity.

Critical weaknesses
  1. The mathematical notation and symbol usage throughout the paper are insufficiently rigorous and consistent, which substantially affects its readability. In particular, numerous subscripts and superscripts are introduced without clear definitions, and several mathematical expressions employ unconventional or ambiguous notation. The authors should follow established mathematical notation conventions, such as those reflected in the math_commands.tex provided by the ICLR LaTeX template.

  2. In my view, the problem investigated in this paper is closely related to personalized federated learning (FL). However, the authors appear to deliberately distinguish their setting from personalized FL without sufficiently justifying this distinction. More importantly, the experimental evaluation overlooks several relevant state-of-the-art personalized FL methods, particularly those addressing heterogeneous tasks, modalities, or label spaces, such as Factorized-FL [1], Co-LoRA [2], FD-Fed [3], and the work by Breitholtz et al. [4]. These studies were identified through a preliminary literature search, suggesting that the current related work coverage and baseline selection may be incomplete. More comprehensive comparisons with representative personalized FL methods are necessary.

  3. The privacy protection mechanisms appear insufficiently investigated. In particular, word-frequency statistics, including the occurrence count of each word, may themselves reveal sensitive information about local datasets and therefore should not be directly uploaded to the server without appropriate protection. Furthermore, it remains unclear whether the shared representations or embeddings are vulnerable to embedding inversion attacks, which could potentially reconstruct sensitive information from the original data. A more comprehensive privacy analysis, including evaluations against representative embedding inversion attacks, is necessary to substantiate the privacy protection claims.

  4. The experimental evaluation is limited to relatively small and simple datasets and models, raising concerns about the scalability and generalizability of the proposed method. In particular, the current experiments do not sufficiently demonstrate its applicability to more complex learning scenarios. Additional evaluations on larger datasets, non-classification tasks, and more advanced model architectures, such as Transformers and potentially large language models, are required.

References

[1] Factorized-FL: Personalized Federated Learning with Parameter Factorization & Similarity Matching. NeurIPS 2022.

[2] Co-LoRA: Collaborative Model Personalization on Heterogeneous Multi-Modal Clients. ICLR 2026.

[3] Feature-driven Layer Specialization for Label Heterogeneous Federated Learning. Neurocomputing 2026.

[4] Federated Learning with Heterogeneous and Private Label Sets. ECML PKDD 2025.

Questions for authors

Please see the above weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.