acceptodds
Under review as a conference paper at ICLR 2027

Graph Federated Unlearning for Privacy Preservation

Abstract

Graph federated learning (GFL) facilitates decentralized training on distributed graph data while keeping sensitive user information local, aligning with policies such as GDPR and CCPA that grant users the right to freely join or withdraw from learning systems. However, even decentralized, user information can persist after quitting, potentially propagating to central servers and then redistributing to malicious clients. This privacy leakage during user withdrawal, despite its importance, has received seldom attention in GFL. To fill the gap, we explore the potential of machine unlearning (MU) to thoroughly remove user information. However, classical MU methods are known to degrade overall performance, a problem that is exacerbated in GFL due to local message passing and global model collaboration. To this end, we make two adjustments to mitigate this challenge for GFL. First, we ensure unlearning updates that minimally affect overall performance, steering them in directions orthogonal to the gradients from learning other data. Second, we introduce virtual clients, maintained by the central server, to preserve graph topology and global embeddings without recovering information of removed entities. We conduct comprehensive experiments under a representative user-withdrawal scenario and propose a novel membership inference framework to rigorously evaluate and validate the reliability of our privacy preservation. The experimental results demonstrate the effectiveness of our approach, which also surpasses the performance of seven state-of-the-art baseline methods.

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

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_q4ZVOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies client-level machine unlearning in graph federated learning. After a client withdraws and requests deletion of its data, its influence may remain in the shared model. The authors propose FedGCV, which consists of two main modules: first, gradient correction adjusts the unlearning update of the target client so as to minimize harm to the performance on retained data; second, the virtual client uses spectral information, feature statistics, and a VGAE to generate synthetic subgraphs to compensate for the loss of graph topology and model performance after deleting the client. Experiments use accuracy and MIA rate as metrics, compare with multiple federated unlearning baselines, and show that FedGCV achieves the best utility–unlearning trade-off among non-retraining baselines.

Critical strengths

The paper addresses a problem with clear research value. Since GNNs aggregate information from neighboring nodes through message passing, client-level deletion may affect the representations and predictive performance of retained clients. Therefore, compared with deleting independent samples in standard federated learning, unlearning in graph federated learning is subject to more complex structural dependencies. The two modules proposed by the authors correspond to two issues in the unlearning process: gradient correction is designed to reduce the harm caused by the unlearning update to the performance of retained clients, while the virtual client aims to compensate for the loss of graph structure and model utility caused by deleting the client subgraph. Under the four datasets and evaluation protocol reported in the paper, FedGCV achieves the highest accuracy and the lowest MIA rate among non-retraining methods, indicating a favorable utility–unlearning trade-off in this experimental setting.

Critical weaknesses
  1. The evidence provided for the paper's stronger claims about privacy preservation and thorough unlearning is still insufficient. Virtual-client artifacts may retain information from the client to be forgotten, limiting the interpretation to model-level empirical forgetting if these artifacts are retained. Moreover, the reported MIA results do not establish privacy against attacks on these artifacts. In addition, the implementation settings explicitly describe optimization against a fixed pre-unlearning loss threshold, while MAIT reuses the same threshold for evaluation. Therefore, the lower MIA rate may partly reflect optimization tailored to this specific attack protocol and does not by itself substantiate the stronger claim of thorough unlearning.
  2. The core idea of gradient correction is conceptually related to existing projection-based or conflict-aware federated unlearning methods. It is suggested to further clarify its methodological differences and graph-specific contribution. Although Figure 1 provides component ablations, these do not establish the added value of the proposed correction over existing alternatives with the virtual-client module held fixed. This leaves the additional contribution of the gradient-correction mechanism insufficiently established.
Questions for authors
  1. What is the deletion scope of the virtual-client artifacts? Specifically, are the spectral information, feature statistics, VGAE-derived information, and synthetic subgraphs generated from the client to be forgotten considered part of the deletion request? If they are retained, how should the paper's claims about privacy preservation and thorough unlearning be interpreted?
  2. How does the proposed gradient-correction method differ from existing projection-based federated unlearning methods? In particular, what graph-specific property requires the proposed correction, and how much of the reported improvement comes from gradient correction itself rather than from combining an existing unlearning update with the virtual-client module?
2 comments

We sincerely thank Reviewer q4ZV for the insightful comments and for acknowledging the importance of addressing structural dependencies in GFU. To address your concerns, we have scoped all claims in the revised manuscript to model-level empirical unlearning, added threshold-free Loss-AUC and λm=0\lambda_m = 0 evaluations, and conducted controlled ablations comparing our gradient correction against FedOSD-proj and FedCARE-proj with the virtual client held fixed.

W1 & Q1: The evidence provided for the paper's stronger claims about privacy preservation and thorough unlearning is still insufficient, and MAIT reuses the same fixed pre-unlearning loss threshold used during optimization. What is the deletion scope of the virtual-client artifacts? Specifically, are the spectral information, feature statistics, VGAE-derived information, and synthetic subgraphs generated from the client to be forgotten considered part of the deletion request? If they are retained, how should the paper's claims about privacy preservation and thorough unlearning be interpreted?

We have revised the manuscript throughout (Abstract, Sections 1, 3.2, 4.2, and 5) to state explicitly that FedGCV targets model-level empirical unlearning under a bounded client handoff. Target-derived artifacts (Δu\Delta_u and (Asyn,Xsyn,Ysyn)(\mathbf A_{\text{syn}}, \mathbf X_{\text{syn}}, \mathbf Y_{\text{syn}})) are used transiently on the server for Rv=5R_v = 5 recovery rounds and purged immediately afterward. If a deletion request forbids any derived artifacts, they are discarded immediately and recovery relies solely on retained clients.

To verify that low MAIT does not merely reflect optimization against the shared cutoff τpre\tau_{\text{pre}} (λm=3.0\lambda_m = 3.0), we evaluate both λm=0\lambda_m = 0 and an independently calibrated, threshold-free Loss-AUC MIA on Cora and CiteSeer:

VariantCora Acc ↑ / MAIT ↓ / Loss-AUCCiteSeer Acc ↑ / MAIT ↓ / Loss-AUC
Retrain68.67% / 3.70% / 50.21%61.83% / 0.00% / 49.72%
FedGCV (λm=0\lambda_m = 0)80.81% / 7.92% / 53.57%67.53% / 10.46% / 52.48%
FedGCV (λm=3.0\lambda_m = 3.0)79.41% / 5.56% / 53.82%69.07% / 7.94% / 52.14%

Across both datasets, FedGCV (λm=0\lambda_m = 0) already reduces MAIT to 7.92%–10.46%, and FedGCV's threshold-free Loss-AUC (52.14%–53.82%) stays close to the ~50% random-guess baseline of Retrain (49.72%–50.21%), confirming genuine empirical forgetting independent of τpre\tau_{\text{pre}}.

W2 & Q2: How does the proposed gradient-correction method differ from existing projection-based federated unlearning methods? In particular, what graph-specific property requires the proposed correction, and how much of the reported improvement comes from gradient correction itself rather than from combining an existing unlearning update with the virtual-client module?

FedGCV differs from existing projection methods in two ways: (1) It projects onto a single aggregate half-space {d:⟨d,Δr⟩≥0}\{\mathbf d : \langle \mathbf d, \Delta_r \rangle \ge 0\} rather than enforcing strict per-client orthogonality (which over-constrains non-IID graphs as in FedOSD). This preserves a larger feasible descent region so unlearning updates are not overly attenuated by conflicting client directions—preventing incomplete forgetting (e.g., FedOSD's 29.63%–29.84% MIA on Cora/CiteSeer)—while keeping server projection O(d)\mathcal{O}(d) and independent of client count MM. (2) It obtains the conflict reference directly from retained clients' GNN updates over observed neighborhoods rather than server-synthesized Euclidean pseudo-samples (FedCARE). This directly captures multi-hop message-passing dependencies rather than isolated out-of-manifold pseudo-samples, preserving retained graph utility while reusing existing FedAvg updates at zero extra generation or communication cost.

In Figure 1, the w/o GRU (without gradient correction) baseline already keeps the virtual client enabled, showing that gradient correction reduces Cora MIA from 24.07% to 5.56%. To strictly isolate the projection rule itself, we hold the virtual-client recovery module fixed and compare alternative correction rules below:

Correction Rule (+ Fixed VC)Cora Acc ↑ / MIA ↓CiteSeer Acc ↑ / MIA ↓
FedOSD-proj + VC69.78% / 27.78%55.36% / 17.65%
FedCARE-proj + VC70.97% / 66.67%58.74% / 15.87%
FedGCV (Ours)79.41% / 5.56%69.07% / 7.94%

With VC fixed, replacing our gradient correction with FedOSD-proj or FedCARE-proj drops accuracy by 8.4–13.7% and leaves higher MIA, isolating the direct contribution of FedGCV's neighborhood-aggregated half-space projection.

We hope these revisions and controlled experiments clearly define the artifact deletion scope, confirm that forgetting is independent of the fixed cutoff τpre\tau_{\text{pre}}, and isolate the contribution of our gradient correction module. Please let us know if any points warrant further clarification.

Thank you for the detailed response. The proposed experiments and analyses address several of my concerns. I will retain my current rating.

Reviewer_FsNfOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors have proposed a federated graph unlearning framework to remove the influence of withdrawing clients along with the preservation of global model performance. It has two main parts: (1) gradient correction that basically projects unlearning updates to minimize interference with remaining clients and (2) virtual-client generation that can use a variational graph autoencoder and graph Laplacian spectral information to generate synthetic graphs for recovering model utility. Experiments are performed on four datasets to demonstrate improvement in accuracy and reduced attack rates compared with basline unlearning methods. But the framework provides just a local theoretical justification for utility preservation without any formal guarantees of complete unlearning or privacy.

Critical strengths
  1. Paper solves an important problem in federated graph learning and it is well motivated.
  2. The integration of VGAE-based generation with Laplacian spectral projection is an interesting approach.
Critical weaknesses
  1. The techniques used are already well established such as gradient projection, VGAE graph generation, and Laplacian spectral filtering. Their combination can be useful, but the paper does not identify the methodological advances over existing gradient-conflict mitigation and graph-generation approaches. The novelty is limited method wise.

  2. The paper do not tell how the withdrawn clients data or the derived statistics are accessed then transmitted/kept and removed during the unlearning process. These things are important for knowing whether the method can support realistic requests or not.

  3. The virtual-client approach use a forgotten clients Laplacian eigenvectors and VGAE representations. But these quantities require the structural information from original client graph. This creates an issue. If the server can already obtain these representations, what information has been transmitted outside the withdrawn client? If the client must compute and provide them during unlearning, what happens when the client is no longer there? The proposed architecture needs a clearly defined sequence of information. Otherwise, it is difficult to assess whether the framework is actually compatible with its motivating federated privacy setting.

Questions for authors

Refer Weaknesses

2 comments

We sincerely thank Reviewer FsNf for the positive assessment of our problem motivation and spectral-guided virtual-client design, as well as the constructive feedback on architectural clarity. In the revised manuscript, we have sharpened our methodological distinctions over existing projection and generation techniques and added an explicit four-stage protocol detailing data access, transmission, post-recovery purge, and offline-client fallbacks.

W1: The techniques used are already well established such as gradient projection, VGAE graph generation, and Laplacian spectral filtering. Their combination can be useful, but the paper does not identify the methodological advances over existing gradient-conflict mitigation and graph-generation approaches. The novelty is limited method wise.

FedGCV's contribution is tailored to the message-passing structure of federated GNNs rather than combining generic modules in isolation:

  1. Graph-native conflict reference: Instead of per-client orthogonalization (FedOSD) or Euclidean pseudo-samples (FedCARE), Eq. (10) directly uses Δr\Delta_r aggregated from retained clients' actual graph neighborhoods, where message passing couples node features and topology.
  2. Spectrally constrained structural recovery: Projecting VGAE latents onto the low-frequency Laplacian subspace (UkUk⊤Z\mathbf U_k \mathbf U_k^\top \mathbf Z) constrains the decoded adjacency to preserve global connectivity for subsequent GNN message passing. Removing this virtual client drops model accuracy on Cora from 79.41% to 65.37%.

W2 & W3: The paper does not tell how the withdrawn client's data or derived statistics are accessed, transmitted, kept, and removed during unlearning. Moreover, Laplacian eigenvectors and VGAE representations require the original graph structure. If the server can already obtain them, what leaked? If the client must provide them, what happens when the client is no longer there? The architecture needs a clearly defined sequence of information.

We have added a concise protocol summary in Section 3.2 (and full specification in Appendix A.4) defining the federated privacy boundary:

  1. Bounded pre-exit handoff: Before disconnecting, the withdrawing client jj locally computes Δu\Delta_u, the Laplacian eigenvectors Uk\mathbf U_k, and the VGAE synthetic subgraph (Asyn,Xsyn,Ysyn)(\mathbf A_{\text{syn}}, \mathbf X_{\text{syn}}, \mathbf Y_{\text{syn}}). Only Δu\Delta_u and the synthetic subgraph are uploaded; raw data (Aj,Xj)(\mathbf A_j, \mathbf X_j), Uk\mathbf U_k, and VGAE parameters never leave client jj.
  2. Server-side correction & recovery: The server executes Eq. (10) using Δu\Delta_u and Δr\Delta_r, then conducts Rv=5R_v = 5 recovery rounds with retained clients and the virtual client.
  3. Post-recovery purge: Upon completion of recovery, the server deletes (Asyn,Xsyn,Ysyn)(\mathbf A_{\text{syn}}, \mathbf X_{\text{syn}}, \mathbf Y_{\text{syn}}) and temporary unlearning updates.
  4. Offline-client fallback: If client jj disconnects after sending Δu\Delta_u, the server applies Eq. (10) and performs recovery using retained clients only. If it disconnects before sending Δu\Delta_u, retraining on retained clients is used (or retained-only fine-tuning when zero derived artifacts are permitted).

We hope this explicit specification of the information flow, client availability handling, and graph-specific methodological advances resolves your concerns regarding federated privacy boundaries and soundness. Thank you again for your valuable suggestions.

Thank you for the detailed rebuttal and clarifications. I am satisfied with the authors' responses, which have addressed my main concerns. I appreciate the additional efforts and will retain my positive rating.

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

This paper proposes FedGCV, a graph federated unlearning framework that aims to remove the influence of a withdrawn client while preserving model utility. It introduces gradient correction to reduce interference with retained clients and VGAE-based virtual clients to compensate for lost structural and feature information. The method is evaluated on four graph datasets using model accuracy and membership inference metrics, with comparisons against existing unlearning methods and component-level ablations.

Critical strengths
  1. The paper addresses a meaningful problem in graph federated learning. Combining gradient correction with virtual-client-based utility recovery provides an intuitive approach to balancing forgetting and performance preservation. The first-order analysis also offers some justification for the gradient correction mechanism.

  2. The experimental evaluation covers multiple unlearning baselines, including retraining from scratch, and reports both model utility and membership inference results. The ablations and additional experiments with different GNN architectures and client counts provide useful evidence of the method's empirical effectiveness.

Critical weaknesses
  1. The novelty of the gradient correction mechanism is limited. FedCARE (IJCAI 2026) employs a nearly identical constrained projection to mitigate interference with retained knowledge, while FedOSD (AAAI 2025) has explored gradient orthogonalization in federated unlearning. FedGCV mainly adapts this idea to aggregated FedAvg updates, introducing the approximation gr∝−Δrg_r\propto-\Delta_r without sufficiently justifying its validity. The lack of discussion and comparisons with these closely related methods further weakens the novelty claim.

  2. The virtual-client mechanism lacks sufficient justification and implementation clarity. Gaussian feature sampling does not explicitly preserve feature–structure correlations, while the supervision and aggregation procedures remain unclear. Moreover, constructing virtual clients from withdrawn-client information raises privacy concerns, as acknowledged by the authors. These issues make it difficult to fully assess the effectiveness and reliability of the proposed recovery mechanism.

Questions for authors
  1. How reliable is gr∝−Δrg_r\propto-\Delta_r under multi-step local optimization, and why use model updates rather than directly aggregated gradients (e.g., FedSGD)? Please also clarify the relationship between aua_u and Δu\Delta_u and provide a more complete derivation of Eqs. (7)–(10). More importantly, the proposed gradient correction closely resembles FedOSD (AAAI 2025) [1] and especially FedCARE (IJCAI 2026) [2], whose constrained projection formulation is nearly identical to Eqs. (7)–(8). What is the technical novelty beyond using FedAvg updates as gradient proxies, and how does the method compare empirically with these approaches?

  2. How does virtual-client generation preserve feature–structure correlations? How are synthetic labels, training objectives, and aggregation weights determined? What privacy assumptions permit retaining information derived from withdrawn clients?

  3. How are α\alpha and τpre\tau_{\mathrm{pre}} determined? Does γ\gamma in Eq. (15) correspond to τa=0.7\tau_a=0.7? Sensitivity analyses and an ablation of fixed-threshold alignment would improve experimental transparency.

  4. Beyond MIA, could the authors consider complementary evaluations, such as prediction agreement with retraining or graph-specific information leakage? These could provide a more comprehensive assessment of unlearning effectiveness.

Overall, I find the problem meaningful and the proposed framework coherent, with encouraging empirical results. However, my main reservation concerns the methodological novelty, particularly the substantial overlap between the gradient correction mechanism and existing projection-based federated unlearning methods. The remaining questions about virtual-client construction and experimental reproducibility further limit my confidence in the contribution. I would welcome further discussion with the authors and would be open to raising my rating if these concerns are convincingly addressed.

[1] Pan, Zibin, et al. "Federated unlearning with gradient descent and conflict mitigation." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 39. No. 19. 2025.

[2] Yue Li, Mingmin Chu, Xilei Yang, Da Xiao, Ziqi Xu, Wei Shao, Qipeng Song, and Hui Li. 2026. FedCARE: Federated Unlearning with Conflict-Aware Projection and Relearning-Resistant Recovery. In Proceedings of the Thirty-Fifth International Joint Conference on Artificial Intelligence, IJCAI-26. IJCAI, 5576–5584. https://doi.org/10.24963/ijcai.2026/621

2 comments

We sincerely thank Reviewer qZEc for the constructive comments and positive assessment of our framework. We have clarified the multi-step FedAvg approximation bound and Eqs. (7)–(10) derivations, detailed the virtual-client design and privacy scope, and provided empirical comparisons against FedOSD and FedCARE alongside the λm=0\lambda_m = 0 ablation, threshold-free Loss-AUC, and retrain-agreement evaluations.

Q1 (Part 1): How reliable is gr∝−Δr\mathbf{g}_r \propto -\Delta_r under multi-step local optimization, and why use model updates rather than directly aggregated gradients (e.g., FedSGD)?

We adopt FedAvg instead of FedSGD because multi-step local training captures graph message passing while reducing communication rounds. From a shared checkpoint w(t)\mathbf w^{(t)}, a single full-gradient step with shared step size η\eta and consistent aggregation weights yields −Δr/η=gr-\Delta_r / \eta = \mathbf g_r exactly. Under KlocK_{\text{loc}} local steps with LL-Lipschitz local gradients, −Δr/(Klocη)-\Delta_r / (K_{\text{loc}}\eta) equals a weighted average of gradients along the clients' local trajectories, whose deviation from gr\mathbf g_r is bounded by O(L∑i≠jpˉimax⁡s∥wi,s(t)−w(t)∥2)\mathcal{O}(L \sum_{i \neq j} \bar p_i \max_s \Vert \mathbf w_{i,s}^{(t)} - \mathbf w^{(t)} \Vert_2). Reusing Δr\Delta_r avoids an additional single-point gradient evaluation and upload from every retained client. We have summarized this in Section 3.1 with the full bound in Appendix A.3.2.

Q1 (Part 2): Please also clarify the relationship between au\mathbf{a}_u and Δu\Delta_u and provide a more complete derivation of Eqs. (7)–(10).

au=∇L^u\mathbf{a}_u = \nabla \hat{\mathcal{L}}_u is the single-point target-loss gradient in the local derivation, whereas Δu\Delta_u is the multi-step parameter displacement returned by the NPO optimizer; therefore Δu=ηuau\Delta_u = \eta_u \mathbf{a}_u holds only for single-step vanilla gradient ascent. In gradient space, minimizing 12∥d−au∥22\frac{1}{2}\|\mathbf{d} - \mathbf{a}_u\|_2^2 subject to ⟨gr,d⟩≤0\langle \mathbf{g}_r, \mathbf{d} \rangle \le 0 yields the KKT solution in Eq. (8): a~u=au−max⁡{⟨au,gr⟩,0}∥gr∥22gr\widetilde{\mathbf{a}}_u = \mathbf{a}_u - \frac{\max\{\langle \mathbf{a}_u, \mathbf{g}_r \rangle, 0\}}{\|\mathbf{g}_r\|_2^2}\mathbf{g}_r (with a~u=au\widetilde{\mathbf{a}}_u = \mathbf{a}_u when gr=0\mathbf{g}_r = 0). Substituting the directional approximation gr∝−Δr\mathbf{g}_r \propto -\Delta_r maps the feasible half-space to ⟨d,Δr⟩≥0\langle \mathbf{d}, \Delta_r \rangle \ge 0, yielding the update-space projection and server rule in Eq. (10), namely: Δ~u=Δu−min⁡{⟨Δu,Δr⟩,0}∥Δr∥22Δr,w(t+1)=w(t)+Δr+βΔ~u.\widetilde{\Delta}_u = \Delta_u - \frac{\min\{\langle \Delta_u, \Delta_r \rangle, 0\}}{\|\Delta_r\|_2^2}\Delta_r, \qquad \mathbf{w}^{(t+1)} = \mathbf{w}^{(t)} + \Delta_r + \beta \widetilde{\Delta}_u.

Q1 (Part 3): More importantly, the proposed gradient correction closely resembles FedOSD (AAAI 2025) and especially FedCARE (IJCAI 2026), whose constrained projection formulation is nearly identical to Eqs. (7)–(8). What is the technical novelty beyond using FedAvg updates as gradient proxies, and how does the method compare empirically with these approaches?

While Eqs. (7)–(8) share the half-space projection form of FedCARE, FedGCV makes two graph-specific advances:

  1. Reference source: FedOSD enforces strict orthogonality to each retained client (over-constraining non-IID graphs), and FedCARE estimates reference gradients from server-generated class-conditional pseudo-samples. FedGCV directly reuses Δr\Delta_r computed on retained clients' observed graph neighborhoods with zero server-side synthesis. Doing so ensures that the conflict reference faithfully captures real topology-coupled message passing—avoiding the structural and distribution mismatch of Euclidean pseudo-samples that leads to severe utility collapse (26.98%–41.30% Acc for FedCARE)—while incurring zero additional server-side generation overhead or communication rounds.
  2. Topological recovery: FedGCV projects VGAE latents onto the low-frequency Laplacian subspace, so the decoded adjacency preserves global connectivity for subsequent federated GNN message passing.

Under the setup identical to Table 1, FedGCV achieves the strongest utility–unlearning balance across all four datasets, whereas FedOSD retains high membership leakage (8.65%–29.84% MIA) and FedCARE collapses utility on citation graphs (26.98%–41.30% Acc) or fails to unlearn on Tolokers (38.41% MIA):

MethodCora Acc ↑ / MIA ↓CiteSeer Acc ↑ / MIA ↓PubMed Acc ↑ / MIA ↓Tolokers Acc ↑ / MIA ↓
FedOSD (AAAI 2025)57.40% / 29.63%70.48% / 29.84%79.45% / 13.72%67.10% / 8.65%
FedCARE (IJCAI 2026)26.98% / 0.00%34.23% / 7.94%41.30% / 21.01%67.95% / 38.41%
FedGCV (Ours)79.41% / 5.56%69.07% / 7.94%83.35% / 3.66%69.17% / 0.18%

Following our first reply to Reviewer qZEc, we address Q2–Q4 below.

Q2: How does virtual-client generation preserve feature–structure correlations? How are synthetic labels, training objectives, and aggregation weights determined? What privacy assumptions permit retaining information derived from withdrawn clients?

  1. Feature–structure correlation: The VGAE encoder jointly processes (Xj,Aj)(\mathbf X_j, \mathbf A_j) to infer latent Z\mathbf Z, which is projected onto the low-frequency Laplacian subspace UkUk⊤Z\mathbf U_k \mathbf U_k^\top \mathbf Z before inner-product edge decoding. This aligns synthesized edges with the global structural skeleton while sampling features from the inferred distribution, enabling GNN message passing to couple both during recovery.
  2. Objectives, labels, and α\alpha: VGAE is trained with standard binary cross-entropy reconstruction plus KL regularization. Synthetic labels Ysyn\mathbf Y_{\text{syn}} are sampled from the withdrawn client's class prior. During recovery, the virtual client minimizes cross-entropy on (Asyn,Xsyn,Ysyn)(\mathbf A_{\text{syn}}, \mathbf X_{\text{syn}}, \mathbf Y_{\text{syn}}) and its update is aggregated with weight α\alpha.
  3. Privacy scope: We scope FedGCV to model-level empirical forgetting under a bounded pre-exit handoff (Section 3.2, Appendix A.3.3, and Appendix A.4), where derived statistics are retained only during the RvR_v recovery rounds and purged afterward.

Q3: How are α\alpha and τpre\tau_{\text{pre}} determined? Does γ\gamma in Eq. (15) correspond to τa=0.7\tau_a = 0.7? Sensitivity analyses and an ablation of fixed-threshold alignment would improve experimental transparency.

α\alpha is chosen via grid search on a validation split from retained clients (searching {0.05,0.1,0.2,0.5}\{0.05, 0.1, 0.2, 0.5\}). τpre\tau_{\text{pre}} is calibrated on the pre-unlearning model θpre\boldsymbol{\theta}^{\text{pre}} as the per-sample cross-entropy threshold minimizing classification error between target members (Dtgt\mathcal D_{\text{tgt}}) and retained-derived non-members (Dvir\mathcal D_{\text{vir}}). Yes, γ\gamma in Eq. (15) corresponds to τa=0.7\tau_a = 0.7 in Section 4.1; we have unified both to τa\tau_a in the revised manuscript. Below we report the ablation removing fixed-threshold alignment (λm=0\lambda_m = 0):

Alignment Variant (Cora)Acc ↑MAIT ↓Loss-AUC (ideal ≈\approx 50%)
FedGCV (λm=0\lambda_m = 0)80.81%7.92%53.57%
FedGCV (λm=3.0\lambda_m = 3.0)79.41%5.56%53.82%

Even with λm=0\lambda_m = 0, FedGCV lowers MAIT to 7.92% with slightly higher accuracy (80.81%), and its threshold-free Loss-AUC remains nearly identical (53.57% vs. 53.82%), confirming that unlearning does not rely on overfitting τpre\tau_{\text{pre}}. We will include the α\alpha-sensitivity curve in the Appendix.

Q4: Beyond MIA, could the authors consider complementary evaluations, such as prediction agreement with retraining or graph-specific information leakage? These could provide a more comprehensive assessment of unlearning effectiveness.

We have added two complementary evaluations in the revised manuscript: (1) Retrain prediction agreement (test-set label agreement with wretrain\mathbf w_{\text{retrain}}) to assess functional fidelity; and (2) Threshold-free Loss-AUC MIA (independent of τpre\tau_{\text{pre}}) to assess privacy without threshold coupling. Regarding graph-specific leakage, (Asyn,Xsyn)(\mathbf A_{\text{syn}}, \mathbf X_{\text{syn}}) are reconstructed solely from the truncated low-frequency subspace Uk\mathbf U_k and perturbed feature moments (σx=0.1\sigma_x = 0.1), and are purged after Rv=5R_v = 5 rounds; we will include quantitative structural/feature reconstruction checks in the final version:

Metric (Cora / CiteSeer)Retrain Agreement ↑Loss-AUC (vs. Retrain)
FedGCV (Ours)85.83% / 82.65%53.82% (50.21%) / 52.14% (49.72%)

We hope these clarifications, derivations, and new empirical results thoroughly address your concerns—particularly regarding the distinctions from FedOSD and FedCARE as well as virtual-client transparency. We welcome any further questions you may have.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.