acceptodds
Under review as a conference paper at ICLR 2027

Deletion Must Reach Every Copy: Federated Execution of Unlearning

Abstract

Federated unlearning (FU) removes the influence of withdrawn data from a model trained by clients that never centralize it. FU faces a distinctive obstacle: federated training leaves persistent copies on devices and at the edge, and each copy an edit misses returns the withdrawn content through later aggregation. Current methods edit the server's global model, and they certify that model rather than the copies a request governs. To address these, we first establish unlearning as an operation on the federated state: the server transcript does not determine which parties a request affects, and holder-local readings do. We then design LOCUS, a Locality-Ordered Correction of Unlearned State framework that executes each step where its evidence lives. Within this framework, Influence Localization has each holder report the retention cost of a candidate correction and what its copy still carries. Correction Propagation then delivers one minimum-norm correction to every holder in that closure and gates a copy out of aggregation until it acknowledges. In addition, Divided Verification assigns removal, retention and application checks to the parties that observe them and binds them into one certificate. Across two benchmarks, three deployments, five unlearning objectives and seven published FU protocols, LOCUS corrects 92–99% of the governed copies, while server-side execution corrects none and the protocols at most 60%. For example, on TOFU with Llama-3.2-1B, LOCUS with NPO corrects 95% of the copies and preserves the erasure throughout a 300-round continuation, whereas server-side NPO loses it after 17 rounds and Oblivionis, the longest-surviving published protocol, after 73 rounds.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
federated unlearning, personalized federated learning, machine unlearning, verification, data deletion, distributed systems

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_EHvhOpenReview ↗
Rating
3(+1)
Confidence
3(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies federated unlearning when the model has persistent copies, such as personalized adapters, offline clients and released checkpoints. The authors argue that current methods only edit the server's global model, so uncorrected copies keep the withdrawn content and can bring it back through later aggregation. Thus, they propose LOCUS, which has three stages. (1) Influence Localization has each holder report local readings to decide which copies need correction. (2) Correction Propagation sends one minimum-norm correction to all of them and blocks uploads from holders that haven't acknowledged it. (3) Divided Verification combines the removal, retention and acknowledgment checks into one certificate. Experiments on TOFU and MUSE-News with 20 clients show that LOCUS corrects 92% to 99% of the affected copies and keeps the erasure through 300 rounds of continued training, while server-side unlearning corrects none.

Critical strengths
  1. The problem is important and under-explored. The paper makes a good point that federated training leaves copies everywhere, and that editing only the server model misses them. Figure 2 shows that uncorrected copies can bring the forgotten content back within a few rounds. Table 9 suggests that very few existing protocols deal with this.

  2. The idea of running each step where its evidence lives is intuitive. The acknowledgment gate is simple but effective, and the ablation in Table 2 shows that each of the three stages matters.

  3. The experiments are extensive. The paper covers five unlearning objectives, seven federated unlearning protocols, three deployments and a 300-round continuation. It also includes broadcast and retraining controls, sequential requests, adversarial reflow, and different model families and client counts.

  4. The paper is open about its limits. It reports how much the localization step leaks (Table 7) and where forged acknowledgments break the certificate (Table 4).

  5. The figures and tables are clear and nicely designed. At a time when many papers feel rushed, I can see the care the authors put into their figures and tables, and I really learned a lot from them.

Critical weaknesses
  1. I am worried that a simple control may nearly match LOCUS. Broadcast-all with the same gate reaches full coverage on all three deployments and also survives 300 rounds, using 40 messages against 41 for LOCUS (Tables 1 and 5). The main gain of LOCUS is lower retention excess and slightly higher utility. Meanwhile, the headline numbers mainly compare against methods that were never designed to touch the copies, and server-side coverage is zero by construction. So the value of localization may be smaller than the abstract suggests.

  2. The method relies on retrained references. The thresholds γc\gamma_c and the retention baselines come from matched leave-requester-out retraining for every copy class. However, the paper doesn't explain how they are obtained in practice, and their cost isn't counted in Table 5.

  3. The theory does not fully support the claims. Theorem 4 is a linear extrapolation that assumes a positive descent rate at every round. Theorems 8 and 11 show that the server can't know what it never receives, which is rather intuitive. None of the results gives a formal unlearning guarantee for the corrected copies, and the certificate is essentially a set of measured loss thresholds.

  4. The experimental setup is relatively small. By default, the paper uses 20 clients, a single request and rank-16 adapters, and each 7B deployment is run only once.

  5. This paper introduces too many new terms (e.g., carriage reading, participation closure, governed copies, reflow, release margin), and the abstract is hard to follow on a first read. I suggest the authors polish the writing in a later version.

Questions for authors

I would be thankful and willing to raise my score if the authors can (partly) address the following questions/concerns.

  1. Is the benefit over Broadcast-all with the gate only in retention excess? Could the authors show a setting where broadcasting to everyone causes a clear utility loss and localization really matters. For example, with many more clients or a larger retention cost?

  2. How are γc\gamma_c and the retention baselines obtained without retraining? How do results change when the thresholds are estimated from a proxy instead of a matched reference?

  3. What happens with a dishonest client that forges its acknowledgment or sends restoring updates after acknowledging?

  4. Can the copy probe be made mandatory, and what does it cost?

  5. Could the authors report results with more clients?

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

This paper addresses a critical gap in Federated Unlearning (FU): the failure of existing methods to account for persistent model copies on client devices. The authors argue that uncorrected copies reintroduce withdrawn data through aggregation reflow in subsequent training rounds. They propose LOCUS, a framework that executes unlearning where the evidence lives. Experiments show LOCUS corrects 92%–99% of copies and maintains erasure over 300 continuation rounds, significantly outperforming server-side execution and seven published FU protocols.

Critical strengths
  • The reframing is consequential and carries a clean negative result. Treating the request as governing the federated state (§2, Eq. 1) makes coverage of persistent copies the binding constraint, which Figure 2 and Observation 1 quantify; more importantly, Table 4 shows that three server-grounded certificates and the single-party certificates of VeriFi and BlockFUL hold 0.0% sound pairs, i.e. they necessarily accept an unapplied copy, whereas the divided certificate reaches 93.8%. That finding is worth reporting on its own.

  • The empirical scope and the reporting discipline are strong: three deployments, five objectives, seven protocols, a stage ablation, five certificate states and overhead accounting, with variance for the 1B runs, censored survival times recorded as such, and retention excess and utility reported alongside coverage.

Critical weaknesses
  • I was unable to locate the boundary between this paper and two closely related works. Lethe (arXiv:2601.22601, Jan 2026) already names the failure mode in which continued training on retained clients revives unlearned knowledge, and proposes a method against it; this is close to Observation 1 and Figure 2(b). veriFUL (IEEE Internet Computing 2025, arXiv:2510.00833) already formalizes the verification entities, goals, approaches and metrics for verifiable federated unlearning, which is close to §3.4 at the framework level.

  • The first line of the certificate appears to certify a different quantity from the one the paper uses to decide whether a copy still carries the withdrawn content. Eq. (7) uses a loss threshold, Lf(wc′)≥γcL_f(w'_c)\ge\gamma_c, whereas §2 and Figure 2 measure the share of Df\mathcal{D}_f that a copy still answers correctly; γc\gamma_c and η\eta are fixed by the deployer from matched retrained references, with no distributional bound. Table 7 measures what the localization readouts release, not whether a corrected copy still permits recovery of Df\mathcal{D}_f. If the gap is real, relating min⁡c(Lf(wc′)−γc)\min_c\big(L_f(w'_c)-\gamma_c\big) to the success of a subsequent extraction probe would make the word "corrected" considerably firmer.

Questions for authors

Q1: Table 2 lowers survival from ≥ 300 to 212–220 when DV is disabled although §4.3 presents DV as verification only, so could you define what "DV off" removes and separate the verification from the release decision in the ablation?

Q2: Since closure is the union of the gcg_c and mcm_c sets (Eq. 2) but Table 2 ablates IL only as a whole, could you add the gcg_c-only and mcm_c-only rows against the same C column?

Q3: Given that Lethe opens with the same motivation about persistence under continued training and that veriFUL already formalizes verifiable federated unlearning at the framework level, could you state explicitly in §1 and Appendix A how this work differs from the two, neither of which appears in the references?

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

The paper proposes LOCUS, a federated unlearning framework that aims to remove the influence of the withdrawn content from persistent model copies stored by each client in the federation. LOCUS consists of three steps. Starting from the candidate direction δ\delta, Step 1 (Influence Localization) determines participation closure P\mathcal{P} (i.e., set of clients that either have high forget loss or have local gradient similar to δ\delta) and optimizes δ\delta in the minimum-norm problem. Step 2 (Correction Propagation) propagates the optimized δ\delta to clients in P\mathcal{P} and gates their updates from subsequent server aggregation until the clients send back their acknowledgement. Step 3 (Divided Verification) certifies unlearning completion based on acknowledgement messages and forget and utility evaluation results.

Critical strengths
  1. LOCUS is rigorously designed to remove influence of withdrawn data from persistent copies. All components are well-motivated. Additionally, the paper theoretically justifies information reflow from persistent model copies, bounding the number of training rounds to recover withdrawn data post-unlearning.
  2. LOCUS shows strong empirical effectiveness, significantly reducing number of training rounds to recover withdrawn data across different datasets (TOFU, MUSE), model scales (1B-7B), and model families (Llama, Qwen).
Critical weaknesses
  1. The problem solved by LOCUS is not well-justified. In LOCUS, clients preserve persistent model copies and send back their updates by locally fine-tuning from these persistent copies. Differently, clients in typical FL receive the broadcasted model from the server and start their local fine-tuning from the broadcasted model. Hence, when unlearning is performed on the server (broadcasted) model in typical FL, it is expected that the information about withdrawn data should not be introduced back to the model during later aggregation.

  2. The paper writing should be improved. The paper uses many uncommon and inconsistent technical jargons that are not widely used in typical FL/FL unlearning papers and not sufficiently explained in the paper, such as "aggregation reflow", "carriage", "participation closure", "coverage", "transcripts", "holders"/"parties" (how are they different from "clients"?), "minimum copy margin". This makes certain parts of the papers difficult to comprehend (e.g., section Methodology, Figures 3, Table 1) and overall less attractive to read.

Questions for authors
  1. How realistic is the problem solved by LOCUS (information reflow due to local fine-tuning on persistent copies)? Does this problem arise because clients can be dishonest (i.e., deliberately use their persistent copies instead of the server broadcasted model for local fine-tuning)? If clients can be dishonest, does this affect how influence localization and divided verification are designed (in influence localization, an affected client could send incorrect gcg_c and mcm_c (in Eq. 2) to best preserve its utility; in divided certification, an affected client could incorrectly flag the unlearning update as violating the retain performance)?

  2. What algorithm is used to solve the minimum-norm problem in Eq. 3? How can we ensure there is a solution Δc\Delta^c that satisfies all the conditions? If there is no feasible Δc\Delta^c, how would your method behave?

  3. Relating to the previous question, why do we only find a single Δc\Delta^c for all clients in the closure instead of different Δc\Delta^c for each client? Intuitively, each persistent model copy held by each client can be affected differently by the withdrawn data, which should motivate different Δc\Delta^c.

  4. On Line 261, it is mentioned that the requester would release a probe set drawn from DfD_f and other clients would evaluate their own copies on this probe set. Do you think this conflicts to the standard FL goal, where clients only share their updates but not their data?

  5. (Clarifying question) Could you clarify why κ\kappa defined as such in Eq. 4 as well as its underlying assumption?

  6. (Clarifying question) How do you obtain γc\gamma_c in Eq. 2? It is said that γc\gamma_c refer to the forget loss of the retraining reference. However, this violates unlearning assumption that retraining server model is undesirable due to its expensiveness.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.