acceptodds
Under review as a conference paper at ICLR 2027

FedRUn: Efficient Client-Side Federated Unlearning with Rényi Entropy

Abstract

With the rise of privacy regulations, machine learning models should be able to remove the influence of designated data from the trained model upon request. In Federated Learning (FL), this task becomes more challenging because the training data remain distributed across multiple clients while only model updates are shared with the server. Retraining without target data provides a natural reference for unlearning, but requires substantial distributed computation and communication costs. Existing federated unlearning (FU) methods may help to reduce this cost, but most of them rely on stored training history, previous global models, or recovery procedures involving non-requesting clients, which incurs excessive runtime, massive historical storage, repeated communications, and intensive computation overhead. To address these limitations, we propose FedRUn, an efficient client-side FU framework based on Rényi entropy maximization. FedRUn performs unlearning directly from the final trained global model without requiring stored historical model updates or repeated communication and computation costs, reducing the system overhead and making unlearning very efficient. The unlearning process is performed locally at the target client, eliminating the need for the server to store historical client model updates while allowing each client to keep its data locally. Since the target client has direct access to its local data, FedRUn can apply the Rényi-entropy objective directly to either a subset of samples or the entire local dataset. Specifically, FedRUn maximizes predictive Rényi entropy on the forget set to reduce predictive confidence, while retaining and proximal terms preserve non-forgotten samples and limit model changes. Only the resulting client-computed unlearning update is sent to the server to update the global model. This design enables client-level, sample-level, and multi-client unlearning within the same framework while reducing computation, communication, and storage requirements.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Federated Unlearning, Federated Learning, Machine Unlearning, Client-Side Unlearning, Rényi Entropy, Sample-Level Unlearning, Multi-Client Unlearning, Membership Inference.

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_VqFVOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2(+1)
Presentation
2(−1)
Soundness
2
Read review
Paper summary

The paper proposes FedRUn, a federated unlearning framework that maximizes the entropy on forget data at requesting clients. Clients return locally computed updates for server-side application, avoiding historical-update storage and additional participation from non-requesting clients.

Critical strengths

Strengths:

  1. The method cover a wide range of unlearning type, including the client-, sample-, and class-level unlearning.
  2. The method show a knowledgeable improvement in terms of computation and communication efficiency.
Critical weaknesses

Weaknesses:

  1. The literature of the paper from recent years seems to be limited, thus, limiting the novelty and significance claim of the paper and the research questions. For instance, in the introduction, the paper claimed that current methods still depend on historical states, repeated recovery rounds, or method-specific assumptions. However, there are recent works in 2025-2026 (e.g., R1, R2, R3, R4, R5) already tackles these issues. A better statements and literature should be made and discussed to improve the paper significance. Also, as the baselines seem to be old, it would be good if the authors consider some literature from recent years.

     - R1: Unlearning through Knowledge Overwriting: Reversible Federated Unlearning via Selective Sparse Adapter, CVPR 2025.
    
     - R2: NoT: Federated Unlearning via Weight Negation, CVPR 2025. 
    
     - R3: Computation and Communication Efficient Federated Unlearning via On-server Gradient Conflict Mitigation and Expression, CVPR 2026.
    
     - R4: Efficient and Effective Federated Unlearning via Adaptive Data Partitioning, TKDE 2026.
    
     - R5: Forgettable federated linear learning with certified data unlearning, TNNLS 2026.
     
    
  2. The contribution of maximize the entropy of the prediction on the TA, FA in tables 1, 8, 9 seem not to be significant over other recent works.

  3. Many parts in the paper lack consistency and coherence.

  4. The method seems to be incremental. The entropy maximization for the forgetting loss seems not to be a novel contribution. Many methods, especially series of gradient ascent, have used it widely [R6, R7, R8]. Meanwhile, the proximal loss appear in FedProx [R9].

     - R6: A Closer Look at Machine Unlearning for Large Language Models, ICLR 2025.
    
     - R7: Certifying the Right to Be Forgotten: Primal–Dual Optimization for Sample and Label Unlearning in Vertical Federated Learning, IEEE TIFS.
    
     - R8: Maximizing Local Entropy Where It Matters: Prefix-Aware Localized LLM Unlearning, ACL 2026.
    
     - R9: Federated Optimization in Heterogeneous Networks
    
Questions for authors
  1. What do you mean "these methods still depend on historical states, repeated recovery rounds, or method-specific assumptions"?
  2. Why Fast-FedUL contribute no unlearning computation on the client side, no storage cost and no communication cost? Why do the reported communication cost in Fast-FedUL in table 11 is reported with 446.69 MB, and in table 12, the Fast-FedUL is reported with 10710 MB?
Reviewer_vBnaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper have proposed FedRUn, a client-side framework for post-training federated unlearning. FedRUn can perform client-side federated unlearning through Rényi entropy maximization. It usually requires one communication exchange and no stored training history, supporting several unlearning granularities. The framework could support client-level, sample-level, and also simultaneous multi-client unlearning. Experiments cover multiple datasets and IID/non-IID settings which showing substantial cost reductions but varying utility and forgetting performance.

Critical strengths
  1. The method FedRUn obviously reduces computation, communication, and the need to store training history.
  2. Evaluation of experiments covers multiple unlearning granularities and both IID and non-IID settings.
Critical weaknesses
  1. It is unclear whether Rényi entropy offers an advantage over Shannon entropy or uniform-target losses, since the paper does not compare them under the same conditions. Why does Rényi entropy work in this case?
  2. The baselines are from 2021–2024 which are too old. More recent relevant methods, for example, FedOSD (AAAI 2025), are cited but not included or compared in the experiments.
  3. Some details needed to reproduce the results are missing, including a code link, the membership inference attack (MIA) setup, baseline settings, and the hardware and software environment. This makes it hard to independently check the reported forgetting performance and efficiency gains.
  4. The main experiments use only 5–10 clients. The paper does not vary the amount of training data or the number of clients, so it is unclear how well the method scales.
Questions for authors
  1. What advantage does Rényi entropy provide over Shannon entropy and uniform-target losses mentioned in the paper? Could you provide further evidence?
  2. Could you compare against recent methods such as FedOSD, or explain why they were left out?
  3. Could you provide more details about the MIA setup and provide stronger forgetting validation and scaling experiments?
Reviewer_yctBOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FedRUn, an efficient client-side federated unlearning framework based on Rényi entropy maximization. The method performs post-training unlearning locally at the requesting client without requiring historical model updates or additional participation from non-requesting clients. Specifically, FedRUn maximizes predictive Rényi entropy on the forget set to reduce model confidence, while incorporating retain and proximal regularization terms to preserve utility and control parameter deviations when applicable. The resulting local update is transmitted to the server to construct the unlearned global model. The authors evaluate FedRUn across multiple datasets, including MNIST, CIFAR-10, Purchase-10, and CIFAR-100, considering client-level, sample-level, and multi-client unlearning scenarios. The experimental results demonstrate substantial reductions in computation, communication, and historical storage requirements compared with retraining and several existing federated unlearning methods, while maintaining competitive predictive utility and reducing membership-inference exposure.

Critical strengths

A major strength of this work is its practical system design. FedRUn performs unlearning directly from the final trained global model and requires only the requesting client to participate in local optimization. This avoids storing extensive training histories and eliminates the need for repeated recovery rounds involving other clients. The efficiency improvements are substantial. For example, on CIFAR-10, FedRUn completes unlearning in 1.58 seconds with 89.44 MB of communication, compared with 874.6 seconds and 7157.47 MB for full retraining. These results demonstrate the practical value of the proposed framework in resource-constrained federated learning environments.

The paper evaluates FedRUn using multiple representative baselines and considers several unlearning granularities, including client-level, sample-level, multi-client, and class-level unlearning. The additional experiments on non-IID data distributions and hyperparameter sensitivity further strengthen the empirical evaluation. In particular, the results on MNIST and CIFAR-10 suggest that FedRUn can maintain predictive utility close to retraining while considerably reducing system overhead. Overall, the experimental coverage provides meaningful support for the practical applicability of the proposed framework.

Critical weaknesses

Although Rényi entropy maximization is central to the proposed method, the paper does not provide a direct comparison with alternative entropy-based forgetting objectives, such as Shannon entropy maximization, under identical experimental settings. While the sensitivity analysis of the Rényi order is useful, such a comparison would help establish the specific benefit of the selected objective and strengthen the methodological contribution.

The paper primarily evaluates forgetting effectiveness through forget-set accuracy and membership-inference metrics. Although these experiments provide useful evidence of reduced membership exposure, they do not fully characterize the extent to which the influence of forgotten data has been removed. Additional evaluation using complementary unlearning metrics or stronger membership-inference attacks would provide a more complete assessment, especially for cases where differences from retraining remain noticeable.

Questions for authors

Q1. Could the authors clarify whether Rényi entropy maximization provides measurable benefits over standard Shannon entropy maximization under comparable computational budgets? A controlled comparison would help establish the specific contribution of the proposed forgetting objective.

Q2. Could the authors further clarify the extent to which the reported membership-inference results reflect successful unlearning rather than primarily reduced prediction confidence? Additional discussion or complementary evaluation would strengthen the interpretation of the results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.