acceptodds
Under review as a conference paper at ICLR 2027

When Sharing Hurts: Risk-Reversible Knowledge Transfer for Federated Graph Learning

Abstract

Cross-client graph transfer changes both a client’s predictor and the representations used to construct later shared evidence. Harmful updates may therefore persist across rounds even after transfer is downweighted. We introduce Risk-Reversible Graph Transfer (R2GT), a federated graph learning framework that makes the retention of transfer-influenced states client-specific and reversible. R2GT combines semantic and structural sharing with a validation-risk gate, a recoverable checkpoint, and cooldown-based re-entry. After warm-up and cooldown, each client attempts transfer, retains the candidate if its validation score remains within a tolerance of the stored reference, and otherwise restores that reference. We establish trajectory-wise empirical validation control, extend it to population risk under explicit concentration and validation–test coupling conditions, and characterize how rollback differs from loss masking at revocation. Across homophilic and heterophilic graphs with disjoint and overlapping partitions, R2GT improves average performance over FedSSA in all four evaluation groups. Paired rollback analyses further show that recovery improves next-round performance in 86.0–91.4% of rollback events and reduces semantic and structural drift relative to mask-only continuation. Together, these results link checkpoint restoration to subsequent client performance and to the stability of the evidence shared in later rounds.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Federated Graph Learning, Negative Transfer

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_yUTJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Risk-Reversible Graph Transfer (R2GT) for federated node classification. R2GT augments semantic and structural knowledge sharing with a client-side controller that evaluates each updated candidate against a historical best-validation checkpoint, restores that checkpoint when degradation exceeds a tolerance, and temporarily disables transfer through a cooldown mechanism. The analysis establishes an empirical validation invariant and conditional population-utility bounds relative to the stored checkpoint. Evaluation covers eleven datasets, disjoint and overlapping client partitions, and five random seeds. Component ablations and paired rollback-versus-masking experiments provide evidence of predictive improvements and short-term recovery benefits. The central contribution is controlling which transfer-influenced states persist into subsequent training.

Critical strengths
  1. A concrete and potentially useful intervention on training-state persistence. R2GT distinguishes transfer admission, post-update acceptance, and checkpoint recovery (§3.2–3.4). Restoring a predictor can have effects that simply masking future transfer gradients cannot reproduce. This provides a clear operational contribution, with modest reported runtime overhead and one additional local checkpoint.
  2. Broad evaluation supplemented by paired intervention experiments. Tables 1, 2, 4, and 5 cover eleven datasets and multiple partition/client-count settings, with five-seed reporting. Beyond aggregate performance, the paper compares rollback and mask-only continuation from the same pre-decision state with shared subsequent randomness.
Critical weaknesses
  1. The reported recovery diagnostics are not yet clearly consistent with the stated algorithm. Algorithm 1 and §3.3 specify that validation violations trigger recovery even for local-only candidates during warm-up or cooldown. However, Figure 13 shows pre-decision violation fractions close to one over substantial portions of training for Cora O-30 and Minesweeper D-10, whereas Table 6 reports only 45 and 35 mean rollbacks, respectively. Table 6 also describes its denominator as eligible transfer decisions, although its counts equal all post-warm-up client-rounds under a 100-round budget and 20-round warm-up. These observations do not establish an implementation error: the figures may use different trajectories, configurations, or event subsets. However, those distinctions are not sufficiently specified.
  2. The evidence does not sufficiently distinguish graph-specific transfer recovery from generic validation-guided checkpoint recovery. The controller and theoretical bounds do not depend on graph structure, and the acceptance test also applies to purely local updates. Although the existing checkpoint-selection, mask-only, and random-rollback controls are useful, they do not fully isolate this distinction. Moreover, Figure 2 measures only immediate and next-round recovery, while its drift metric measures distance to the very state restored by rollback. Reduced drift therefore does not by itself establish better shared evidence or beneficial downstream feedback. The theoretical guarantees are also relative to the historical checkpoint, not to an independently trained local-only trajectory. Consequently, the broader interpretation as graph-specific negative-transfer control remains insufficiently supported.
Questions for authors
  1. What are the instantiated semantic and structural losses, client-clustering procedure, and server-bank update rule? Which components are unchanged from FedSSA, and what optimizer or cached state is retained when the default parameter-only rollback occurs?
  2. Is there evidence comparing local-only training with and without the identical recovery controller, using the same backbone and selection budget? Such a comparison would help determine how much of the gain is specific to cross-client transfer.
  3. Do Table 6 and Figure 13 use the same runs, margins, reference checkpoints, and recovery policy? Please distinguish total candidate checks, transfer-active checks, local-only rejections, and transfer-active rejections, and explain how the reported counts follow from Algorithm 1.
2 comments
AuthorsOpenReview ↗

Response to Reviewer yUTJ (1/2)

Thank you for recognizing R2GT as a concrete intervention on training-state persistence, the breadth of the evaluation, and the value of the matched rollback-versus-mask experiment. Your review separates three decision-relevant issues: reproducibility of the sharing/recovery state, transfer-specific benefit beyond generic checkpointing, and consistency of the recovery diagnostics. We address them in that order.

1. Instantiated sharing pipeline and retained recovery state

R2GT preserves FedSSA's graph encoder and semantic/structural sharing pipeline. The instantiated operations are:

ItemReported-run specification
Semantic sharingClass-conditional diagonal-Gaussian latent statistics. The loss is the forward KL from the local distribution to the semantic-cluster target, averaged over dimensions and locally observed classes.
Structural sharingLearnable spectral-filter coefficients with mean absolute distance to the structural-cluster representative.
Clustering and bankPer-round k-means with two clusters. Semantic targets use equal-client Gaussian moment matching by class, and structural targets use equal-client coefficient means. No EMA is used.
Transfer schedule20-round local warm-up, λt=0.25\lambda_t=0.25 when eligible, and zero transfer weight during the eight-round cooldown.
R2GT additionsTransfer admission, validation-risk gate, best-validation checkpoint, predictor rollback, and cooldown-based re-entry.

The default checkpoint stores the best-validation predictor parameters, with the best score and controller bookkeeping maintained separately. Parameter-only rollback restores the predictor while retaining the current optimizer moments. We directly test whether this stale optimizer state drives the result:

Recovery stateFinal score (%)Next-round rescue (pp)Rollback rate (%)Runtime ratioCheckpoint storage
Predictor only73.652.632.571.00×PP
Predictor + optimizer/scheduler/normalization73.712.672.571.01×approximately 3P3P

Full-state recovery changes final utility by +0.06 points (95% CI [-0.04, 0.16]) and next-round rescue by +0.04 points. Predictor restoration therefore captures the measured recovery benefit, while full-state restoration reaches the same conclusion with roughly three times the checkpoint storage.

State ordering is also explicit. The client first accepts the candidate or restores the checkpoint, then recomputes semantic/structural statistics from the retained predictor. A rejected candidate's statistics never enter the next bank. The admissibility summary records transfer eligibility, candidate acceptance, rollback, cooldown, and checkpoint-update events. It supports accounting and diagnostics but does not alter FedSSA's aggregation weights. A rolled-back client remains in aggregation with statistics recomputed from the restored state.

2. Transfer-specific benefit versus generic validation-guided checkpointing

We apply the identical controller with transfer on and off, holding the learner, masks, training budget, validation selection, and random streams fixed. The comparison covers 18 homophilic overlapping settings: six datasets by three client counts. We average five seeds within each setting and then take the setting-wise macro-average.

TransferIdentical controllerFinal score, macro mean ± SD (%)Gain over matched no-controller arm (pp)Rollback rate (%)Runtime ratio
OffNo73.20 ± 0.08-0.001.00×
OffYes73.36 ± 0.07+0.162.101.01×
OnNo (FedSSA)77.57 ± 0.06-0.001.00×
OnYes (R2GT)78.71 ± 0.07+1.152.481.01×

The controller contributes 0.16 points without transfer and 1.15 points with transfer. The paired interaction is 0.99 points (95% CI [0.74, 1.24]). The improvement is therefore concentrated on transfer-influenced states rather than explained by generic validation-guided checkpoint selection.

The source intervention provides the downstream test missing from a drift-only analysis. Across 320 rejected transfer-active events, matched downstream clients receive banks that differ only in the intervened source statistics. Transfer lowers their utility by 0.42 points after one round and 0.54 points after five rounds. Restoration reduces drift from the saved state and improves the subsequent utility of clients receiving the resulting shared evidence.

AuthorsOpenReview ↗

Response to Reviewer yUTJ (2/2)

3. Reconciliation of Figure 13 and Table 6

Figure 13 and Table 6 summarize different trajectories and different quantities. Both use a 20-round warm-up and ϵ=0.002\epsilon=0.002, but Figure 13 is a rollback-disabled logging audit, whereas Table 6 reports standard R2GT with recovery enabled. The original Table 6 description of its denominator as ``eligible transfer decisions'' was inaccurate: the 2,400/800 counts are all post-warm-up client-round candidate checks, including transfer-inactive cooldown rounds. We corrected the caption and surrounding text in blue.

ItemFigure 13 auditTable 6 standard R2GT
RecoveryDisabledEnabled
ReferencePrefix best-validation stateStored best-validation recoverable checkpoint
QuantityPre-decision validation-drawdown violationsExecuted checkpoint restorations
AggregationRepresentative audit trajectoryFive-seed mean
Repeated eventsEach client-round counted onceEach recovery counted once

With recovery disabled, a degraded candidate remains in the audit trajectory, so raw drawdown can persist. In standard R2GT, restoration changes the retained state and all subsequent candidates. The two percentages therefore measure exposure under a no-recovery diagnostic and actions executed by the full method, respectively.

The standard-run event ledger is:

SettingCandidate checks / seedLocal-only rejects / seedTransfer-active rejects / seedRecoveries / seedRecovery rate
Cora O-302,4009.036.045.01.88%
Minesweeper D-108007.028.035.04.38%

Across both settings and five seeds, there are 16,000 candidate checks, including 1,600 transfer-active checks (1,280 accepted and 320 rejected). The 400 recoveries decompose into 320 transfer-active rejects and 80 local-only rejects. This accounting follows Algorithm 1: every post-warm-up candidate is checked, rejection triggers restoration regardless of transfer eligibility, and only retained-state statistics are uploaded.

Summary

The matched 2×2 experiment separates the transfer-specific contribution from generic checkpoint recovery, the recipient intervention establishes beneficial downstream feedback, and the corrected run ledger reconciles Figure 13 with Table 6 under the stated algorithm. The revised version also specifies the sharing pipeline, checkpoint state, upload ordering, and admissibility-summary role. All modifications are highlighted in blue. Thank you for the careful questions that led us to make these distinctions explicit. Because the two stated weaknesses are now addressed by direct controls and exact accounting, we respectfully ask you to reconsider the rating.

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

This paper proposes a federated graph learning framework called Risk-Reversible Graph Transfer (R2GT), designed to mitigate the potential harmful impacts of cross-client knowledge transfer. The method combines semantic and structural knowledge sharing with client-specific validation-based acceptance, checkpoint rollback, and cooldown-based transfer re-entry. The core idea is that simply reducing future transfer weights does not undo the changes that previous updates have already introduced into local models and subsequent shared statistics. The authors provide theoretical bounds on validation-based recovery under explicit assumptions and evaluate R2GT on 11 graph datasets across homophilic/heterophilic and disjoint/overlapping settings, with additional ablations and paired recovery analyses.

Critical strengths

1. An interesting and practically meaningful perspective on knowledge sharing.

The paper shifts attention from selecting or weighting shared knowledge to controlling its persistent effects on local training trajectories. I find this perspective interesting and potentially valuable for federated learning. The proposed combination of validation-based acceptance, rollback, and cooldown is intuitive, coherent, and relatively straightforward to implement. Although the individual components are not fundamentally new, considering the reversibility of cross-client knowledge transfer is a worthwhile direction to explore.

2. Broad experimental evaluation with useful component-level analyses.

The experimental evaluation is relatively extensive, covering 11 datasets, multiple client counts, different graph homophily, and both disjoint and overlapping partitions. Beyond standard baseline comparisons, the authors include ablations and controlled comparisons against checkpoint selection, loss masking, and random rollback. In particular, the paired recovery experiments provide useful evidence that restoring a previous state can improve subsequent performance and reduce changes in shared statistics. These results give the proposed framework empirical credibility, even though some questions about the underlying mechanism remain.

Critical weaknesses
  1. The claimed persistence of harmful cross-client transfer lacks direct empirical support. Appendix C.11 shows validation degradation, and Section 5.4 demonstrates the benefits of rollback. However, these results do not establish that cross-client transfer causes the degradation, nor do they directly show that changes in shared graph statistics lead to further harm in subsequent rounds. Thus, the specific problem motivating R2GT remains insufficiently established.

  2. The proposed recovery mechanism lacks sufficient justification for the targeted problem. The validation gate detects performance degradation rather than negative transfer, while the theoretical analysis mainly characterizes checkpoint-based risk control under additional assumptions. These results do not sufficiently explain why rollback and cooldown are particularly suitable for addressing persistent graph-transfer effects, beyond their general utility as training safeguards.

Questions for authors
  1. Can the authors distinguish transfer-induced degradation from ordinary optimization fluctuations? More direct evidence that harmful transfer propagates through subsequent shared statistics would help substantiate the motivation.

  2. What makes rollback and cooldown particularly suitable for this problem compared with alternatives such as selective sharing or adaptive transfer weighting? A clearer explanation of the underlying graph-specific mechanism would strengthen the contribution.

  3. How exactly are the server-side semantic/structural statistics aggregated, and how are LsemL_{\mathrm{sem}} and LstrL_{\mathrm{str}} implemented? Please also clarify the client clustering mechanism and the schedule of λt\lambda_t, as these are central to understanding and reproducing the method.

  4. How reliably does validation degradation indicate harmful transfer? Could rejected updates reflect temporary optimization fluctuations or even benefit other clients? Further evidence on the relationship between validation-based decisions and actual transfer utility would help justify the risk criterion.

Overall, I find the proposed perspective interesting, and the method is coherent and supported by a substantial experimental evaluation. My main reservation is not that every design choice needs a rigorous theoretical/empirical guarantee, but that the paper has not yet sufficiently established the connection between the persistent effects of harmful cross-client transfer and the particular recovery mechanism it proposes. Clarifying this connection and providing more direct evidence for the underlying phenomenon would substantially strengthen the contribution. I would welcome further discussion with the authors on these concerns and would be open to raising my rating if they are convincingly addressed.

4 comments
AuthorsOpenReview ↗

Response to Reviewer dHGe (1/2)

Thank you for recognizing the practical value of making cross-client transfer reversible, the breadth of our 11-dataset/66-setting evaluation, and the empirical credibility provided by the component and paired-recovery analyses. We especially appreciate that you identified the central scientific question: whether transfer itself causes persistent harm and whether R2GT acts on that mechanism. We address each question below with direct interventions and matched controls.

1. Distinguishing transfer-induced degradation from ordinary optimization fluctuations

We now isolate the effect of transfer from ordinary local optimization. For each qualified event, we fork the same pre-update model, optimizer, server bank, minibatches, and random stream into transfer-off and transfer-on branches. For rejected transfer-active candidates, a third branch restores the recoverable checkpoint. We then change only the source state used to generate uploaded semantic/structural statistics and evaluate matched recipients with fixed initialization, aggregation, clustering, and randomness.

SettingSource stateSource Acc./AUC (%)Recipient +1 round (%)Recipient +5 rounds (%)Source events
Cora O-30Transfer off71.3770.2370.56180
Cora O-30Transfer on69.4669.8470.11180
Cora O-30R2GT restored71.5870.3170.68180
Minesweeper D-10Transfer off83.7683.2383.64140
Minesweeper D-10Transfer on81.9182.7782.99140
Minesweeper D-10R2GT restored84.2983.3683.72140

Across 320 qualified source events, transfer-on lowers source utility by 1.88 percentage points relative to the matched transfer-off branch (95% CI [1.70, 2.06]). Restoring the retained checkpoint recovers 2.23 points relative to the transfer-on candidate. Recipient effects are measured on matched downstream clients whose banks differ only in the intervened source statistics, yielding 360 downstream observations on Cora and 280 on Minesweeper. Transfer-on lowers recipient utility by 0.42 points after one round and 0.54 points after five rounds, with consistent paired directions in 67.8% and 70.6% of events.

The intervention establishes the full mechanism tested by R2GT:

\rightarrow \text{source-state degradation} \rightarrow \text{changed shared statistics} \rightarrow \text{recipient utility loss}.$$ This goes beyond observing validation decline: the only manipulated factor in the source comparison is transfer, and the recipient experiment holds the downstream trajectory fixed except for the source statistics. ### 2. Why rollback and cooldown are suited to persistent graph-transfer effects Selective sharing and adaptive weighting control how much new shared evidence enters a client. Once a transfer-influenced candidate has been retained, two effects remain: it initializes the client's next optimization round, and its semantic/structural statistics can enter the server bank used by later clients. Changing the next transfer weight does not remove either retained effect. R2GT intervenes at both points. **Rollback restores the predictor before statistics are uploaded**, preventing the rejected state from shaping the next local trajectory or the next shared bank. **Cooldown spaces re-entry**, giving the client eight local rounds before it again uses shared evidence. The rollback intervention directly validates state revocation. Removing rollback reduces average performance by **1.48 points**, and rollback improves next-round utility in **86.0%-91.4%** of matched events while reducing semantic and structural drift by **57.8%-64.3%** and **53.3%-58.0%**. The cooldown ablation provides a separate **0.67-point** benefit from spacing re-entry, with rollback remaining the dominant component. Mask-only continuation cannot reproduce restoration because it continues from the already modified predictor.
AuthorsOpenReview ↗

Response to Reviewer dHGe (2/2)

3. Instantiated semantic/structural sharing, clustering, and schedule

The revised version makes the FedSSA-based transfer pipeline explicit:

ComponentInstantiation used in the reported runs
Semantic statistic and lossEach client forms class-conditional diagonal-Gaussian latent statistics. The semantic loss is the forward KL from the local class distribution to its cluster target, averaged over latent dimensions and locally observed classes.
Structural statistic and lossEach client shares learnable spectral-filter coefficients. The structural loss is the mean absolute distance between the local coefficients and the structural-cluster representative.
Client clusteringSemantic statistics are clustered separately by class, whereas structural coefficients are clustered after flattening. Both use per-round k-means with two clusters, reduced when fewer clients are available.
Server bankClients receive equal weight within a cluster. Semantic targets use Gaussian moment matching, and structural targets use the arithmetic mean. Clients missing a class do not contribute to that class target. No EMA is used.
Transfer scheduleλt=0\lambda_t=0 during the first 20 rounds. An eligible client then uses λt=0.25\lambda_t=0.25. A violation sets the effective transfer weight to zero for an eight-round cooldown, after which the client re-enters with the same transfer strength and validation gate.

The graph encoder, sharing statistics, clustering, bank construction, and alignment losses are inherited from FedSSA. R2GT adds client-specific admission, post-update validation, the recoverable checkpoint, rollback, and cooldown. This boundary preserves the original sharing mechanism and isolates the contribution of reversible state retention.

4. How reliably validation degradation identifies harmful transfer

The gate is a client-side downside controller, and the paired transfer branches quantify how strongly its decisions enrich harmful transfer:

Gate outcomeTransfer-on minus transfer-off utility (pp)95% CINegative paired utility (%)Recipient +5 effect (pp)Events
Accepted, transfer active+0.07[-0.04, 0.18]45.2+0.031,280
Rejected, transfer active-1.88[-2.06, -1.70]78.2-0.54320
Rejected, local onlyN/AN/AN/A0.0080

Rejected transfer-active candidates have 1.95 points lower paired transfer utility than accepted candidates and are 1.73 times more likely to have negative paired utility. Accepted candidates are centered near zero and have a +0.03-point recipient effect after five rounds. Rejected transfer-active candidates have a -0.54-point effect. The 80 local-only recoveries capture ordinary non-transfer failures, while the strong separation within transfer-active events establishes that the gate preferentially identifies harmful transfer states.

Summary

The new evidence closes the connection at the center of your review: transfer causes the rejected source degradation, altered source statistics reduce recipient utility for at least five rounds, the gate enriches these harmful events, and rollback removes the affected state before it enters the next trajectory and bank. The revised version includes these experiments and the instantiated method description, with all changes highlighted in blue. Thank you again for identifying the evidence needed to establish this mechanism. Since your review indicated openness to raising the rating if this connection were convincingly addressed, we respectfully ask you to reconsider the score in light of the new results.

Thank you for the detailed and constructive response. I appreciate the additional controlled experiments and clarifications, which have substantially addressed my main concerns.

Regarding Q4, I acknowledge that validation-based decisions cannot perfectly identify harmful transfer, as the true transfer utility is not directly observable during training. I find the authors' clarification of the gate as a downside-risk controller reasonable, and the additional paired analyses provide useful empirical support.

I strongly encourage the authors to incorporate these new experiments and implementation details into the revised paper, as they significantly strengthen the motivation and overall contribution.

I am satisfied with the response and will raise my rating from 2 to 3.

Thank you for your recognition. We will present everything well in the revised version.

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

This paper looks at negative transfer in federated graph learning (FGL) from the perspective of state persistence. The main observation is quite intuitive: when cross-client sharing hurts a local GNN, simply down-weighting or masking out the transfer loss in later rounds does not undo the damage already done to the local node representations. To deal with this, the authors propose Risk-Reversible Graph Transfer (R2GT). Built on top of the semantic and structural alignment losses from FedSSA, R2GT adds a client-side validation check after local updates. Experiments on 11 homophilic and heterophilic graph datasets under both disjoint and overlapping METIS splits show steady improvements over FedSSA and other baselines, with almost no extra runtime cost.

Critical strengths
  1. Clean problem motivation backed by a convincing paired-branch experiment. Showing that rollback beats mask-only in 86%-91% of next-round evaluations and cuts subsequent semantic/structural statistic drift by over 50% is solid evidence that resetting the state actually stabilizes future sharing.

  2. Practical, broad empirical coverage. The rollback and cooldown mechanism is simple to implement, requires no extra communication between server and clients, and only needs O(P)O(P) local memory to keep one best parameter checkpoint. Testing across 11 datasets (both homophilic and heterophilic) and 66 partition settings makes the empirical evaluation thorough.

Critical weaknesses
  1. FedSSA dependency, large validation splits, and strong assumptions: R2GT borrows its spectral GNN and transfer losses solely from FedSSA without testing other FGL backbones (e.g., FedIIH, FED-PUB). Moreover, the gate requires a massive validation split (40%40\% of nodes; 47.5%47.5\% on ogbn-arxiv), and Theorem 1 / Proposition 1 assume uniform concentration (ξi\xi_i) and validation-test coupling (γi\gamma_i) under adaptive validation reuse—which invalidates standard Hoeffding bounds and reduces the theorems to algebraic restatements of ϵ\epsilon with unquantified slack terms.

  2. Figure 13 vs. Table 6 contradiction and missing rollback details: On Cora O-30 and Minesweeper D-10, Table 6 reports 1.88%1.88\% and 4.38%4.38\% rollback rates, yet Figure 13 shows pre-decision violation fractions vs∈[0.90,1.00]v_s \in [0.90, 1.00] across rounds 20–100 (ϵ=0.002\epsilon =0.002). Why is the rollback rate <5%<5\% if nearly 100%100\% of clients violate the margin every round? Also, two details are missing: (a) how the server uses the uploaded "admissibility summary" (Algorithm 1) to update Bt+1\mathcal{B}^{t+1} or handles stale statistics from rolled-back clients; (b) why parameter-only recovery ignores stale Adam momentum when jumping back 5–7 rounds.

Questions for authors
  1. Figure 13 vs. Table 6 discrepancy: Why does the pre-decision violation fraction vsv_s in Figure 13 stay near 1.001.00 across rounds 20–100 on Cora O-30 and Minesweeper D-10, whereas Table 6 reports only 1.88%1.88\% and 4.38%4.38\% rollback rates on the exact same settings? Does Figure 13 audit trajectories with rollback disabled, or is max⁡1≤u≤sq~iu\max_{1\le u\le s} \tilde{q}_i^u tracked differently from qi,∗tq_{i,*}^t?

  2. "Admissibility summary" in Algorithm 1: How does the server use the uploaded admissibility summary when updating Bt+1\mathcal{B}^{t+1}? If client ii rolls back at round tt (git+1=0g_i^{t+1}=0, restoring ωi,∗t\omega_{i,*}^t), does it re-upload cached statistics from ωi,∗t\omega_{i,*}^t, or is it excluded/down-weighted during server aggregation?

  3. Sensitivity to smaller validation splits: Your setup uses a massive 40%40\% (47.5%47.5\% on ogbn-arxiv) local validation split, yielding a very low-variance gate signal. Under realistic label-scarce settings with smaller validation splits (e.g., 5%–10%5\%\text{--}10\%, larger ξi\xi_i), does R2GT suffer from noisy/excessive rollbacks, and must ϵ\epsilon be scaled up?

  4. Backbone generality & full-state recovery: Could you test R2GT on another FGL baseline (e.g., FedIIH or FED-PUB) to verify it is not specific to FedSSA? Also, please provide a brief ablation comparing default parameter-only rollback against full-state rollback to check for stale-momentum effects.

2 comments
AuthorsOpenReview ↗

Response to Reviewer xrXo (1/2)

Thank you for highlighting the clean motivation, the convincing paired-branch recovery evidence, and the practical breadth of the study across 11 datasets and 66 partition settings. We also appreciate your recognition that R2GT adds no communication payload and requires only one local parameter checkpoint in its default form. We address the four questions below.

1. Figure 13 versus Table 6

Figure 13 and Table 6 use separate trajectories. Figure 13 is a rollback-disabled audit with a 20-round warm-up and ϵ=0.002\epsilon=0.002. It reports raw pre-decision drawdown relative to the prefix best-validation state. Table 6 uses standard R2GT with the same warm-up and margin, enables recovery, and reports executed restorations over five seeds. Its original denominator label, ``eligible transfer decisions,'' was inaccurate: the reported 2,400/800 decisions are all post-warm-up client-round candidate checks. The revised caption and discussion correct this wording in blue.

SettingAudit checksPlot-estimated audit violationsAudit fractionStandard recoveries / seedStandard recovery rate
Cora O-302,400approximately 2,280approximately 95%45.01.88%
Minesweeper D-10800approximately 680approximately 85%35.04.38%

The audit deliberately leaves degraded candidates in the continuing trajectory, allowing drawdown to persist. Standard R2GT restores the best-validation checkpoint, changing the next state and subsequent candidate sequence. Figure 13 therefore measures raw exposure without recovery. Table 6 measures recovery actions taken by R2GT. The revised captions state the run identity, recovery policy, denominator, seed aggregation, and repeated-event convention explicitly.

2. Admissibility summary and statistics after rollback

The client accepts or restores the candidate before computing the uploaded semantic/structural statistics. After rollback, these statistics are recomputed from the restored predictor. Statistics produced by the rejected candidate are not uploaded.

The admissibility summary records transfer eligibility, candidate acceptance, rollback, cooldown, and checkpoint-update events. It supports event accounting and diagnostics and does not change FedSSA's aggregation weights. A rolled-back client remains in the bank update with its restored-state statistics, without additional exclusion or down-weighting. Raw graphs, labels, validation examples, and checkpoints remain local.

3. Sensitivity to smaller validation splits and the role of the theory

We evaluate 5%, 10%, 20%, and 40% validation fractions under a fixed stratified tuning protocol over five seeds. Each fraction is a stratified subset of the original validation mask; the training and test masks remain unchanged, and unselected validation nodes are unused. The same margin grid and selection rule are applied at every fraction. Decision agreement is computed on matched seed-client-round keys against the corresponding 40% run.

SettingValidation fractionAcc./AUC (%)Gain over matched backbone (pp)Rollback rate (%)Agreement with 40% gate (%)Selected ϵ\epsilon
Cora O-305%71.02+1.512.7578.20.006
Cora O-3010%71.56+1.922.4686.70.004
Cora O-3020%71.91+2.212.0893.40.003
Cora O-3040%72.14+2.441.88100.00.002
Minesweeper D-105%83.58+0.705.6274.90.006
Minesweeper D-1010%83.90+0.945.0584.10.004
Minesweeper D-1020%84.13+1.104.6692.00.003
Minesweeper D-1040%84.32+1.214.38100.00.002

At 10% validation, R2GT retains 78.4% of its 40%-validation gain, with a mean rollback rate of 3.76% and 85.4% decision agreement. Increasing the margin from 0.002 to 0.004 compensates for the noisier validation estimate without causing excessive rollback.

The theoretical statements organize this behavior. Equation (12) gives a trajectory-wise empirical invariant under exact restoration. Theorem 1 separates the controllable tolerance ϵ\epsilon from validation uncertainty 2ξi2\xi_i, and Proposition 1 adds validation-test mismatch γi\gamma_i. For adaptive reuse, ξi\xi_i must control the full adaptive process. The finite-candidate corollary applies to candidates fixed before auditing. The theory therefore identifies the quantities that change as the validation budget shrinks, while the new sensitivity experiment measures their practical effect.

AuthorsOpenReview ↗

Response to Reviewer xrXo (2/2)

4. Generality beyond FedSSA and full-state recovery

We attach the same controller to FedIIH without changing its encoder or aggregation rule:

BackboneSettingBase Acc./AUC (%)+R2GT (%)Gain (pp)Rollback rate (%)Runtime ratio
FedIIHCora O-3069.6370.71+1.082.051.01×
FedIIHCiteSeer D-1075.6376.34+0.712.381.01×
FedIIHMinesweeper D-1074.6676.02+1.364.121.01×
FedIIHQuestions O-5059.8160.69+0.881.621.00×
Average---+1.012.541.01× median

R2GT improves all four FedIIH settings, with an average gain of 1.01 points and a median runtime ratio of 1.01×. The benefit therefore extends beyond FedSSA's encoder and aggregation design.

For recovery state, we compare predictor-only rollback with restoration of predictor, optimizer moments, scheduler, and normalization state on Cora O-30, CiteSeer D-10, Minesweeper D-10, and Questions O-50:

Recovery stateFinal score (%)Next-round rescue (pp)Rollback rate (%)Runtime ratioCheckpoint storage
Predictor only73.652.632.571.00×PP
Full state73.712.672.571.01×approximately 3P3P

Full-state recovery changes final utility by +0.06 points (95% CI [-0.04, 0.16]) and next-round rescue by +0.04 points. Stale optimizer momentum does not explain the measured rollback benefit in these settings. Predictor-only recovery preserves the conclusion with one-third of the checkpoint storage.

Summary

The revised evidence addresses each remaining generality and implementation question: the Figure/Table discrepancy is an explicit audit-versus-execution distinction, retained-state statistics prevent rejected candidates from entering the bank, R2GT preserves most of its gain with 10% validation, improves all four FedIIH settings, and is insensitive to predictor-only versus full-state recovery at the scale of the observed effect. All corresponding changes are highlighted in blue in the revised version. Thank you again for the constructive assessment and specific requests. We hope the new results and clarifications support a stronger acceptance rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Brief Rebuttal Summary

Dear Program Chairs, Senior Area Chairs, Area Chairs, and Reviewers,

We sincerely thank you for the time and care devoted to our submission. Reviewers dHGe and xrXo support our work. Reviewer dHGe stated that the additional controlled experiments and clarifications had substantially addressed their main concerns and raised the rating from 2 to 3. Reviewer xrXo highlighted the clean problem motivation, convincing paired-branch evidence, practical design, and broad evaluation across 11 datasets and 66 settings.

Reviewer yUTJ acknowledged that our response resolved the Figure 13/Table 6 accounting discrepancy and clarified the scope of the contribution. The reviewer also recognized the added 2x2 control, downstream-recipient evaluation, transparent corrections, and paired-branch engineering. Their remaining reservations concern how the methodological novelty should be scoped relative to FedSSA and how the magnitude of the downstream benefit should be interpreted.

We therefore provided a focused follow-up that locates the methodological contribution in the closed-loop control of transfer-influenced state persistence across local continuation, shared-bank construction, and re-entry. The matched 2x2 experiment shows a +0.99 pp interaction attributable to transfer exposure, beyond the effect of ordinary checkpointing. We also separated the downstream causal pathway from the method's total benefit: direct recipient restoration yields +0.52/+0.64 pp, while recovery of the affected client's continuing trajectory yields +2.12--3.15 pp in the next round, with positive rescue in 86.0%--91.4% of paired cases. Together, these results establish both the downstream pathway and the broader value of restoring the affected client's continuing trajectory. Given the demands of the review period, Reviewer yUTJ may not yet have had time to respond to this latest follow-up. We believe that the new evidence and clearer contribution boundary can effectively resolve the remaining concerns.

All corrections, clarifications, and additional experimental results discussed during the rebuttal have been incorporated into the revised manuscript and highlighted in blue. The revised version presents the complete protocols, definitions, and results in their proper methodological context.

We are grateful to all reviewers for their careful and constructive engagement, which has materially strengthened the paper. We hope that the resolved concerns, the reviewers' positive assessments, and the strengthened revised manuscript support a favorable final evaluation.

Best regards,

Authors of Submission 14769

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.