acceptodds
Under review as a conference paper at ICLR 2027

No More Guessing: a Verifiable Gradient Inversion Attack in Federated Learning

Abstract

Existing gradient inversion attacks in federated learning can fail to disentangle aggregated batches and generally lack an intrinsic mechanism to certify reconstruction success. In vision and language domains, an attacker may rely on human inspection to assess reconstruction plausibility, but this can be costly and may limit attack efficacy and scalability under tight time constraints. For numerical tabular records, however, human inspection provides no comparable means of assessing reconstruction plausibility. We propose a verifiable gradient inversion attack (VGIA) that provides explicit certificates for record isolation and reconstruction. VGIA exploits the geometry of fully connected layers by interpreting activation boundaries as hyperplanes that partition the input space. It then applies an algebraic subspace test to detect when a hyperplane-delimited region contains exactly one record. Once a single record has been certified, our method analytically recovers the input features and uses lightweight optimization to reconstruct targets. Certified regions are then removed from further exploration, allowing VGIA to adaptively allocate subsequent hyperplane queries to unresolved regions. Experiments on tabular and image benchmarks demonstrate exact recovery under large batch sizes in regimes where prior methods fail or cannot assess fidelity. Thanks to its adaptive search strategy, VGIA also requires fewer attack rounds than previous approaches.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
federated learning, privacy, gradient inversion

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_jF5hOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigate the verifiability of gradient inversion attacks (GIAs) in FL and propose the verifiable gradient inversion attack (VGIA). Authors highlight that existing attacks can produce spurious reconstructions without a reliable way to verify their attack results. VGIA exploits the geometry of fully connected layers and introduces a subspace-based verfication to determine whether each sample has been isolated in the hyperplanes. After the samples are verfied, VGIA reconstructs the inputs and adaptively refines another regions. Using various tabular and image datasets, authors demonstrate improved reconstruction accuracy and reduced communication rounds.

Critical strengths
  • This paper approaches the GIAs from a novel perspective of the verifiability from the attacker's point of view.
  • The proposed approach is intutive.
  • The experiments are well structured across various datasets.
Critical weaknesses
  • Among the assumptions used in the theoretical argument, N≤3BN \leq 3B is not satisfied in the main experiment. This greatly affects the completeness of a paper, from theory to experiment.
  • The experiments under defense mechanisms are not sufficiently comprehensive. In Appendix H, authors only provide the evaluations under gradient clipping and Gaussian noise injection. Consideration of various defense mechanisms should be included. Like [1-4], these methods employ different defense strategies, including perturbation, pruning and obfuscation. I wonder that "Can VGIA still keep its verification guarantee under the various client-side defenses?".
  • The proposed approach is largely built upon CTP, and it seems that the main novelty of this paper lies in replacing the heuristic criterion with a subspace-based verfication. Although I thought this is a meaningful improvement, the overall attack framework remains somewhat incremental.

[1] Sun, J., Li, A., Wang, B., Yang, H., Li, H., & Chen, Y. (2021, June). Soteria: Provable defense against privacy leakage in federated learning from representation perspective. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 9307-9315). IEEE.

[2] Xue, L., Hu, S., Zhao, R., Zhang, L. Y., Hu, S., Sun, L., & Yao, D. (2024, March). Revisiting gradient pruning: A dual realization for defending against gradient attacks. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 38, No. 6, pp. 6404-6412).

[3] Zhang, K., Cheng, S., Shen, G., Ribeiro, B., An, S., Chen, P. Y., ... & Li, N. (2025). Censor: Defense against gradient inversion via orthogonal subspace Bayesian sampling. arXiv preprint arXiv:2501.15718.

[4] Luo, C., Yau, D. K., & Song, Q. (2025). Svdefense: Effective defense against gradient inversion attacks via singular value decomposition. arXiv preprint arXiv:2510.03319.

Questions for authors

See the weaknesses.

8 comments
AuthorsOpenReview ↗

Introduction and answer to Weakness 1

We thank the reviewer for recognizing the novelty of verifiability from the attacker's perspective and the well-structured experimental evaluation. Below, we address the reviewer's concerns by clarifying the scope of our theoretical assumptions (W1), examining the suggested defenses (W2), and highlighting VGIA's contributions beyond CTP (W3).

W1: N≥3BN\geq 3B condition

The condition N≥3BN \geq 3B is introduced in the “Round Complexity” paragraph (line 350) solely to establish a high-probability upper bound on the number of attack rounds required to certify all target samples. Importantly, this condition is not required for Proposition 4.1, our main theoretical result, which establishes the soundness of the isolation certificate independently of this condition.

As intuition suggests and our round-complexity bound confirms, for a fixed number of target samples BB, increasing the number of attacked neurons NN reduces the number of rounds required by the attack. Our main experiments deliberately consider a more challenging regime (N=1000N=1000, B=2048B=2048), in which the number of attacked neurons is less than half the number of target samples, well outside the scope of the theoretical bound. Nevertheless, VGIA successfully certifies all target samples within just 12–13 rounds (Figure 3).

We will update the paper to further stress that N≥3BN\geq 3B is only required to derive the round complexity bound and that in the experiments we consider a particularly challenging regime.

Finally, we observe that the round-complexity bound can potentially be extended to the regime N<3BN < 3B. The intuition is that the attack can process up to ⌊N/3⌋\lfloor N/3 \rfloor unresolved regions per round, refining the remaining regions in subsequent rounds. This leads to a bound of the same form, with an additional multiplicative factor of order B/NB/N. We do not believe this extension is essential to our main contributions, but we would be happy to provide a formal derivation if the reviewer considers it useful.

AuthorsOpenReview ↗

Answer to Weakness 2 (part 1 of 3)

W2: VGIA guarantee under other defenses

Beyond gradient clipping and Gaussian noise, we evaluated the state-of-the-art client-side detection mechanisms proposed by Carletti et al. (IEEE S&P 2026), as well as an additional, more effective cosine-similarity check that we developed (lines 461–466, Appendix G.4). These detection mechanisms also serve as defenses, since a client that detects an attack can stop participating in the training process.

We emphasize that such systematic evaluations remain uncommon in the literature on malicious-server attacks. While several prior works discuss detectability, and Decepticons (Fowl et al., 2023, Appendix G) experimentally investigates how to evade a simple rank-based inspection, the other relevant attacks—including Robbing the Fed (Fowl et al., 2022), CTP (Diana et al., 2025), Panning (Chu et al., 2023), and LOKI (Zhao et al., 2024)—do not systematically evaluate client-side detection mechanisms. Our evaluation goes further by quantifying detection effectiveness through empirically calibrated false-positive and false-negative rates.

We thank the reviewer for pointing to our attention four additional papers from security venues. Below, we examine how each proposed defense modifies the client's update and discuss the impact of each of these modifications on VGIA’s reconstruction and verification capabilities. We are going to integrate this discussion in the revised version of the paper.

Soteria

Soteria [1] selectively zeros out the gradient components associated with d′<dd' < d coordinates of the input representation to the attacked layer (as confirmed by the official implementation). These components can be easily identified by the attacker. By freezing the parameters of the preceding layers, VGIA can keep the same coordinates masked across attack rounds. The original VGIA algorithm can then be applied without modification to reconstruct and certify the d−d′d-d' unmasked components, provided that the linear-independence assumption of Proposition 4.1 holds for the projected samples in the resulting (d−d′)(d-d')-dimensional space. The server could subsequently modify the preceding layers to change which coordinates remain unmasked, potentially allowing the reconstruction of additional information across rounds. Note that there is an inherent trade-off between privacy protection and model utility: increasing the fraction of masked components d′d' may reduce VGIA's effectiveness, but also degrades training convergence and potentially the final model quality in absence of attacks.

DGP

DGP [2] removes the largest- and smallest-magnitude gradient coordinates in each layer and compensates for omitted updates through error feedback. Since pruning depends on gradient magnitudes, a malicious server could influence which coordinates are retained by appropriately configuring the neurons. In particular, positive ReLU rescaling allows the server to make all nonzero components of a neuron's incoming weight and bias gradients arbitrarily large or small, without changing its activation hyperplane or the network output. The attacker can then make auxiliary neurons absorb the extreme-magnitude pruning, while keeping the gradients of the neurons implementing VGIA's hyperplane queries in the intermediate range retained by DGP. Importantly, DGP's error-feedback mechanism does not compromise VGIA's verification guarantee. Whenever a coordinate is retained, its accumulated residual is reset to zero. Therefore, a coordinate retained in two consecutive rounds contains its exact current gradient in the second round, regardless of its earlier history. Consequently, if VGIA keeps the weight and bias gradients of the attacked neurons unpruned throughout the verification sequence, VGIA original verification guarantee applies under the same assumptions as Proposition 4.1.

AuthorsOpenReview ↗

Answer to Weakness 2 (part 2 of 3)

CENSOR

CENSOR [3] generates layer-wise updates orthogonal to the client's original gradient and selects candidates according to their effect on the local loss. This approach presents an inherent trade-off between privacy protection and optimization. Indeed, for any differentiable convex loss L(w)L(w), an orthogonal update uu satisfies L(w−ηu)≥L(w)L(w-\eta u)\geq L(w), since ⟨∇L(w),u⟩=0\langle\nabla L(w),u\rangle=0. Thus, no orthogonal candidate can decrease the loss. For example, with L(w)=12∥w∥2L(w)=\tfrac12\|w\|^2, every norm-matched orthogonal update satisfies L(w−ηu)=(1+η2)L(w)L(w-\eta u)=(1+\eta^2)L(w), so repeatedly applying such updates would cause divergence. CENSOR's Algorithm 1 avoids this issue by falling back to the original gradient when no orthogonal candidate reduces the loss, but this effectively disables its orthogonality-based privacy protection. This trade-off is also observed empirically: [4, Section VI-B] reports a 14% reduction in final test accuracy when applying CENSOR throughout training on CIFAR-10 with ResNet-18, compared with undefended training. To mitigate this degradation, [4] restricts CENSOR to the initial training epochs, recovering accuracy close to the undefended baseline. However, VGIA can be initiated at any stage of training, so a malicious server could simply launch the attack after CENSOR is deactivated. Overall, applying CENSOR throughout training can incur a substantial cost in model utility, while restricting it to the initial epochs leaves the system just as vulnerable to VGIA once the defense is deactivated.

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

The paper proposes VGIA, a malicious-server gradient inversion attack for federated learning that can certify when a reconstructed record is actually correct, which existing attacks can't do. It treats 1st layer ReLU neurons as hyperplanes that slice the input space. A simple subspace test then checks whether a slice contains exactly one sample, and if so the sample is recovered exactly. The results are strong: all 2048 samples are recovered with no spurious ones in 12 rounds, compared with 20+ for CTP. However, the guarantees depend on fairly idealized assumptions (fed SGD, relu), and the defense experiments are only preliminary

Critical strengths

The paper asks a question that gradient inversion work has mostly ignored: how does the attacker know a reconstruction is actually correct? This matters most for tabular data, where a wrong record looks just as plausible as a right one. I think verifiability is a useful lens for evaluating these attacks

The idea is simple and easy to follow. Checking whether the parent slice vector lies in the span of its sub-slices gives a clean test for isolation, and it removes baseline CTP's need to guess a stopping threshold. The full-batch FedSGD results are convincing: all 2048 records are recovered in 12 rounds with no false outputs, while baseline needs 20 rounds even with the right threshold

Critical weaknesses

The verification guarantee holds only in a narrow, idealized setting. It assumes identical rows in the attacked layer, all-positive downstream weights, full-batch FedSGD, and float64 precision. These choices are likely easy to detect. The stealthier configuration needs many more rounds (90 in the current setup). Under FedAvg, noise, or smooth activations, the certificate becomes a tuned threshold test and wrong reconstructions reappear. These are the settings where verification matters most, so "no more guessing" mainly holds in the easiest case.

VGIA builds on CTP. the main results mostly show that the added certificate and search help, not that VGIA advances the SOTA more broadly. Fishing is cited but not evaluated. For tabular data, TabLeak already estimates how reliable each reconstructed feature is. A comparison of how well each method flags wrong reconstructions would make the verifiability claim much more convincing.

The experiments mainly use a small 3-layer ReLU network. Does it work for LLMs, training methods like RL? The architecture scope seems to be limited

Questions for authors

In a realistic setup, with the stealthy configuration, FedAvg with several local steps, and float32, how many rounds does VGIA need to recover a full batch, and how many of the certified reconstructions are actually correct?

How much of VGIA's advantage comes from the span test itself rather than the adaptive search? And how does it compare with Fishing, which uses the same threat model, and with TabLeak's uncertainty estimate at flagging wrong reconstructions?

10 comments
AuthorsOpenReview ↗

Introduction and answer to W1 (part 1 of 2)

We thank the reviewer for appreciating the novelty of our research question ("a question that gradient inversion work has mostly ignored"), for recognizing verifiability as "a useful lens for evaluating these attacks," and for describing our approach as "simple and easy to follow."

W1, Idealized Settings for the Theoretical Guarantees

We explicitly acknowledge that our formal isolation-verification guarantee holds under the assumptions of Proposition 4.1, including full-batch FedSGD and ReLU activations (Section 5, lines 456–460; Appendix I, lines 1802–1803). This setting is well established in the malicious-server gradient inversion literature (Fowl et al., 2022, 2023; Wen et al., 2022; Chu et al., 2023; Boenisch et al., 2023; Diana et al., 2025). While our formal verification guarantees no longer hold when their underlying assumptions are violated, VGIA's verification test remains fundamentally different from the heuristics employed by the baselines. Indeed, it is derived from a criterion that is provably correct under well-defined assumptions, providing a principled theoretical foundation even in settings where the test becomes heuristic. Our experiments show that this distinction matters in practice: even in settings where formal certification is no longer possible, our approach often achieves reconstruction errors one order of magnitude smaller than those of the baselines CTP or Scale-MIA. Thus, while we acknowledge that the strict "no more guessing" guarantee applies only under the stated assumptions, the underlying verification principle provides substantial empirical benefits even beyond its theoretically guaranteed regime. Below, we discuss how our current submission evaluates VGIA's performance when its theoretical assumptions are not satisfied, and present additional experimental results.

Parameter assumptions.

We clarify that identical attacked-layer rows and positive downstream parameters are convenient sufficient conditions for our theoretical construction, not intrinsic requirements for VGIA to operate. Appendix G develops and experimentally evaluates less detectable alternatives, including Gaussian parameter initialization without requiring all neurons to remain active (G.1), a carrier-neuron construction that restricts malicious modifications to a subset of neurons without enforcing positive downstream weights (G.2), and small directional perturbations that restore full rank while preserving high reconstruction performance (G.3).

Detectability.

We systematically evaluated detectability against the state-of-the-art client-side mechanisms proposed by Carletti et al. (2026) and additional within-round and cross-round cosine-similarity tests that we developed (lines 461--466, Appendix G.4). This allows us to investigate the trade-off between detectability and attack efficiency, largely unexplored in previous malicious-server attacks (Fowl et al., 2022; Boenisch et al., 2023; Zhao et al., 2024; Diana et al., 2025).

Regarding the 90 communication rounds, this result concerns a deliberately constrained configuration using only 100 out of 1,000 neurons to reconstruct batches of 2,048 samples, meaning that the batch is over 20 times larger than the number of attacked neurons. Despite the limited number of attacked neurons, VGIA recovers 99.6–99.8% of the samples within 90 rounds, while maintaining full-rank weights and decreasing training loss. Thus, 90 rounds reflect the trade-off between reduced parameter manipulation and reconstruction speed under unusually large batches, not an intrinsic limitation of VGIA. Smaller batches, usually considered in the literature, will require fewer rounds.

AuthorsOpenReview ↗

Answer to W1 (part 2 of 2)

FedAvg, smooth activations, and noisy updates.

Experiments in our submission consider these aspects and show that VGIA remains substantially more effective than CTP in these settings. Under FedAvg (Appendix I), VGIA consistently outperforms CTP across all four datasets, with dataset-wise minimum improvements ranging from 2.4× to 9.6× in correct reconstructions, while CTP produces at least 6.1–286× as many spurious reconstruction candidates, under the same 30-round budget. Under gradients with additive Gaussian noise (Appendix H), VGIA recovers 76% of the batch. With SiLU and GELU activations (Appendix J), VGIA recovers more than 97% of the samples across all tested settings, with reconstruction errors below 1%, while CTP fails to reconstruct any inputs.
These results show that VGIA's verification mechanism remains effective beyond the setting covered by our theoretical guarantees.

Reduced numerical precision

The main experiments use float64 precision to minimize numerical instabilities and evaluate the proposed theoretical mechanism. Unlike the issues discussed above, however, the impact of reduced numerical precision was not investigated in our original submission. We thank the reviewer for raising this important point, which prompted us to investigate the effect of numerical precision on VGIA's performance. We conducted additional experiments in float32, reproducing the main experimental settings across all five datasets and evaluating different span-check thresholds τ\tau. Each entry reports the percentage of correctly reconstructed samples / spurious reconstructios. Although reduced precision introduces numerical errors, VGIA still recovers between 70% and 93% of the original samples, with relatively stable recovery rates across thresholds spanning two orders of magnitude, confirm that the attack remains effective in float32, while also illustrating the distinction between exact theoretical verification and its finite-precision implementation.

Precisionτ\tauACS Income (%)CIFAR-10 (%)HARUS (%)King County Housing (%)ImageNet (%)
float3210−310^{-3}69.87 / 11.9292.72 / 1.3890.14 / 3.3372.30 / 14.6888.74 / 0.26
float3210−210^{-2}72.44 / 14.6793.07 / 1.7290.30 / 3.5473.10 / 14.4892.12 / 0.21
float3210−110^{-1}72.54 / 14.6690.25 / 3.5887.89 / 4.7973.10 / 14.4888.90 / 1.97
AuthorsOpenReview ↗

Answer to W2/Q2 (part 1 of 2)

W2/Q2: Comparison with CTP, Fishing, and TabLeak

We appreciate the reviewer's recognition that "verifiability is a useful lens for evaluating these attacks." We emphasize that certifying successful reconstruction goes beyond recovering candidate samples, particularly for tabular data, where incorrect reconstructions may appear equally plausible, as also highlighted by the Reviewer. VGIA introduces a new gradient-observation construction enabling formal isolation certification and exploits this information to improve the attack itself. We highlight four contributions that distinguish VGIA from CTP.

Novelty with respect to CTP

We agree that VGIA builds on CTP’s geometric slicing framework, and we appreciate the reviewer’s recognition of the value of verification. Our contribution is to make sample isolation intrinsically verifiable under explicit assumptions, through a gradient-observation construction that supports formal certification and an adaptive search procedure. Below, we highlight four aspects of VGIA that constitute substantive advances beyond CTP.

  1. Formal certification from observed gradients. Under the assumptions of Proposition 4.1, VGIA allows the attacker to certify sample isolation using the observed gradients and known model parameters, without knowing the minimum projected separation ϵw\epsilon_w. CTP’s width-based stopping criterion, by contrast, is sound when ϵ≤ϵw\epsilon\leq\epsilon_w, a condition that the attacker generally cannot establish without prior information about the private batch (lines 205–211 and 422–423). Figure 2c illustrates the failure that can occur when this condition is violated. VGIA therefore provides a checkable certificate where CTP relies on a stopping threshold whose validity depends on unknown data.

  2. A gradient-observation construction that makes verification possible. VGIA’s verification requires changing how gradient observations are generated and compared. CTP’s reconstruction relies on invariant per-sample gradient contributions across rounds. With such invariant coefficients, a parent slice vector is automatically the sum of its sub-slice vectors, so membership in their span would hold even if the sub-slices still contained multiple samples. VGIA instead derives computable slice vectors through rescaling and differencing neuron gradients (lines 251–282), while independently resampling downstream parameters across rounds (Proposition 4.1). This variation is essential to making the subspace test informative under the proposition’s assumptions.

  3. Certification guides the search. Once isolation is certified, VGIA removes the corresponding slices from further exploration and reallocates hyperplane queries to unresolved regions, independently of the certified slices’ widths (Figures 2a–2b, lines 215–235). This avoids spending further rounds refining already isolated samples and allows continued refinement where samples remain unresolved. Certification thus directly determines subsequent attack decisions. Although VGIA requires at least three hyperplanes per explored region (lines 284–286), our experiments show that the savings from avoiding further refinement of certified slices outweigh this additional cost.

  4. Substantial gains even against an oracle-tuned CTP baseline. On both ACS Income and CIFAR10, VGIA recovers and certifies all 2,048 samples within 12 rounds, whereas CTP requires 20 rounds even when supplied with the oracle threshold ϵ=ϵw\epsilon=\epsilon_w: a 40% reduction (lines 427–429). Under smaller budgets, the distinction between recovery and verification is particularly visible: after four rounds on ACS Income, approximately 30% of CTP’s reconstructions are spurious, whereas VGIA produces no spurious reconstructions throughout this experiment (lines 430–434).

The broader contribution is to establish how an attacker can identify successful individual reconstructions without access to ground truth, and use this information to guide subsequent queries. The observation construction, correctness guarantee, and adaptive search form a connected contribution with measurable benefits even against CTP supplied with privileged information about the private batch.

This detailed comparison with CTP also helps address another question raised by the reviewer: "How much of VGIA's advantage comes from the span test itself rather than the adaptive search?"

These two components provide complementary benefits. The span test enables verifiable reconstruction, allowing VGIA to certify sample isolation without knowing the minimum projected separation ϵw\epsilon_w, unlike CTP. The adaptive search exploits this certification to improve efficiency, stopping the exploration of certified slices and reallocating hyperplane queries to unresolved regions. In contrast, CTP cannot identify isolated samples and must continue refining nonempty slices until their widths fall below ϵ\epsilon, even when this threshold is optimally chosen.

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

The paper introduces VGIA, a malicious-server gradient inversion attack that combines hyperplane-based reconstruction with an algebraic test for isolating individual records. Under stated assumptions, the test supports verification and adaptive refinement without prior knowledge of the minimum separation between samples. On ACS Income and CIFAR-10 with 2,048 records, VGIA reports complete recovery within 12 rounds, compared with approximately 20 rounds for CTP using an oracle separation threshold, with no spurious reconstructions in these experiments.

Critical strengths
  1. The paper identifies a meaningful distinction between producing reconstructions and knowing which ones are correct without access to private ground truth. This is particularly relevant for numerical tabular records, where visual plausibility provides little guidance.

  2. The verification mechanism is technically motivated and supported by a conditional theorem. Independently changing downstream parameters changes the coefficients of the samples contributing to each slice. Under the stated local linear-independence and sampling assumptions, the parent vector lying in the span of the refined slice vectors implies singleton isolation almost surely. This supplies an explicit stopping rule that CTP lacks.

  3. The main evaluation directly compares with the closest predecessor, CTP, including an oracle choice of its separation threshold. On ACS Income and CIFAR10 with 2048 records, the paper reports full recovery within 12 rounds versus approximately 20 for oracle-threshold CTP, with no spurious VGIA reconstructions in these experiments. The comparison also exposes the tradeoff: CTP recovers more correct records at some early budgets, while VGIA provides greater confidence in its outputs. Additional datasets and three random seeds support the core result.

  4. The appendices investigate several departures from the exact setting, including FedAvg, smooth activations, clipping, noise, and less conspicuous parameter modifications. The authors explicitly acknowledge that several extensions lose the theorem's guarantees, and that the CNN/Transformer certificate concerns embeddings rather than decoded images.

Critical weaknesses
  1. The practical reliability of the certificate is insufficiently established. Proposition 4.1 assumes local linear independence of private inputs, which the attacker cannot generally verify; Appendix B.1 acknowledges possible false certification when this assumption fails. Moreover, all experiments use double precision and sample-wise gradients, without a clear rank-tolerance sensitivity analysis. The authors should report false-certification rates for dependent or nearly dependent records and assess sensitivity to numerical precision and rank thresholds.

  2. The verification guarantee relies on repeatedly querying the same records. The main experiments use full-batch FedSGD on a fixed local dataset, whereas the FedAvg extension is approximate: Table 10 reports about 256 recovered records alongside 684 spurious candidates for King County Housing. The paper should clearly separate these approximate results from verified recovery and evaluate reliability when batches or participating clients change across rounds.

  3. High recovery and low detectability are not demonstrated under the same attack configuration. The near-complete recovery experiment uses 100 attacked neurons over up to 90 rounds, whereas the single-round detection tests use 10 neurons and the cross-round tests distribute a total budget of 100 neurons across rounds. A matched evaluation should report recovery, false-certification rates, and the probability of detection over the complete attack.

  4. The Scale-MIA comparison does not clearly account for query budgets. VGIA uses repeated adaptive queries, while Scale-MIA is a single-round attack, but Table 11 does not specify VGIA's round budget. Moreover, verified embedding isolation does not guarantee image fidelity: for ResNet, 99.21% isolation corresponds to only 28.13% of decoded images reaching PSNR >= 18. The authors should provide a comparison under matched query budgets or analyze recovery and image fidelity as functions of the number of queries.

Questions for authors
  1. What numerical tolerances implement the main span test, and what are the recovery and false-certification rates under float32 batched gradients and locally dependent or nearly dependent records?

  2. What guarantees or empirical reliability remain when batches or participating clients change across rounds? Can exact and approximate settings be summarized separately, including false-certification rates?

  3. Can the authors provide one matched end-to-end experiment demonstrating high recovery, low false certification, and low trajectory-level detection under the same neuron and communication budgets?

  4. How many client queries support the Scale-MIA comparison, and does VGIA retain its advantage when the comparison explicitly accounts for that budget?

  5. For ResNet and ViT, the certificate concerns isolation of an intermediate embedding, while reconstruction of the original image relies on a learned decoder. Appendix K already reports embedding isolation and image-quality metrics separately, but those quality metrics require ground truth. Can the attacker assess the fidelity of a decoded image without the original image, or does verification stop at the embedding? Please clarify how this limitation qualifies the claim of verifiable reconstruction in these architectures.

5 comments
AuthorsOpenReview ↗

Introduction and answer to Weakness 1 and Question 1

We thank the reviewer for their careful assessment and for recognizing the importance of verifiable reconstruction, the theoretical foundations of our verification mechanism, the thorough experimental comparison with CTP, and our investigation of settings beyond the theoretical assumptions.

Linear independence assumption

As detailed in Appendix B.1, Proposition 4.1 requires linear independence only within each slice, rather than across the entire batch. Since slices typically contain few samples, this is a substantially weaker assumption, and we observed no violations in our experiments (lines 299–302).

The requested recovery and false-certification rates can already be derived from the metrics reported in our experiments (Figures 3a–b–d–e and the appendices). The recovery rate is obtained by dividing the number of correctly reconstructed samples by the batch size. The false-certification rate is the proportion of certified reconstructions that are incorrect. Since VGIA certifies all its output reconstructions, this rate coincides with the proportion of spurious reconstructions among all outputs, which we report as Reconstruction Error (%).

Regarding the request to evaluate dependent or nearly dependent records, the reviewer does not specify how near-dependence should be characterized. We interpret it as strong alignment and spectral concentration among input records. Accordingly, we measured the numerical rank, stable rank (∥X∥F2/∥X∥22)(\|X\|_F^2/\|X\|_2^2), and average pairwise cosine similarity of batches of 2,048 samples across five seeds. Despite numerical ranks below their maximum possible values and low stable ranks (1.16–2.32), VGIA achieves near-complete recovery with negligible spurious reconstructions, showing that strong spectral concentration does not necessarily compromise reconstruction accuracy.

In conclusion, we observed no violations of the within-slice linear independence assumption required by Proposition 4.1 in our experiments. Moreover, although the input records often exhibit strong spectral concentration, this does not prevent VGIA in general from reliably reconstructing them.

DatasetNumerical rank / maximumStable rankAvg. cosine similarity
HARUS431.2 ± 0.4 / 5611.156 ± 0.0010.8241 ± 0.0010
ACS Income379 ± 7 / 5732.317 ± 0.0090.4140 ± 0.0018
CIFAR-101516 ± 5 / 20481.191 ± 0.0020.8140 ± 0.0014
King County Housing17 ± 0 / 181.188 ± 0.0040.8435 ± 0.0023

Numerical tolerances and rank thresholds

We acknowledge that our original submission did not sufficiently explain the numerical implementation of the span-membership test (Algorithm 3, Appendix D.3), briefly mentioned at lines 1761 and 1808. Our implementation uses standard Gram–Schmidt orthogonalization to construct an orthonormal basis QQ of the refined slice vectors. We then compute the projection residual e=s−QQ⊤s\mathbf e=\mathbf{s}-QQ^\top\mathbf{s}, where s\mathbf s is the parent slice vector, and accept span membership when ∥e∥2/∥s∥2<τ\|\mathbf e\|_2/\|\mathbf s\|_2<\tau. The only parameter requiring tuning is the tolerance τ\tau, set to 10−910^{-9} in our experiments. The table below investigates the sensitivity of VGIA's performance to this parameter. We will clarify the numerical implementation and the choice of tolerance in the revised manuscript.

New experiments with batched gradients and reduced precision

As requested, we conducted additional experiments using batched gradients to evaluate the impact of the span-membership threshold τ\tau and numerical precision. The table below reports "correct-certification rate / false-certification rate", averaged over three seeds, for batches of 2,048 samples. The results show that float64 verification is robust across four orders of magnitude of τ\tau, recovering over 96% of samples with a false-certification rate of at most 1.89% across the tested thresholds. While float32 introduces greater numerical errors, it still achieves approximately 90–93% recovery on CIFAR-10, HARUS, and ImageNet, and over 72% on Income and King County Housing, with suitable thresholds. Moreover, float32 remains robust across two orders of magnitude of τ\tau.

Precisionτ\tauACS Income (%)CIFAR-10 (%)HARUS (%)King County Housing (%)ImageNet (%)
float6410−1010^{-10}100/0100/0100/099.95/0.05100/0
float6410−910^{-9}100/0100/099.87/0.0799.95/0.05100/0
float6410−810^{-8}100/099.87/0.0798.97/0.5199.90/0.08100/0
float6410−710^{-7}99.87/0.0799.54/0.2396.16/1.8999.66/0.21100/0
float3210−310^{-3}69.87/11.9292.72/1.3890.14/3.3372.30/14.6888.74/0.26
float3210−210^{-2}72.44/14.6793.07/1.7290.30/3.5473.10/14.4892.12/0.21
float3210−110^{-1}72.54/14.6690.25/3.5887.89/4.7973.10/14.4888.90/1.97
AuthorsOpenReview ↗

Answer to Weakness 2 and Question 2

We understand the reviewer's request as asking us to clarify the conditions under which VGIA's certification guarantees remain valid with FedAvg, intermittent client participation, and varying batches across rounds. We address each of these settings below.

FedAvg

We have explicitly acknowledged in several places that VGIA provides only approximate verification under FedAvg, as the reviewer probably recognizes in Strength 4: "The authors explicitly acknowledge that several extensions lose the theorem's guarantees." Specifically, in Section 4.2 (lines 247–249), we state that verification under FedAvg is approximate; in Section 5 (lines 458–460), we acknowledge its reduced robustness; and in Appendix I (lines 1800–1803), we explain why Proposition 4.1 no longer holds exactly.

Nevertheless, the loss of formal guarantees does not imply a loss of empirical effectiveness. VGIA remains substantially more effective than existing baselines under FedAvg. To our knowledge, no prior malicious-server attack has demonstrated comparable recovery at such large batch sizes without artificially modifying ReLU activations (e.g., using double-bounded activation functions to mitigate hyperplane drift; Fowl et al., 2022; Zhao et al., 2024) or requiring prior knowledge of the client data distribution.

On King County Housing, in the setting highlighted by the reviewer ( B=128B=128, E=1E=1 ), VGIA recovers 4.7× more correct records than CTP (256 vs. 54), while producing 11.8× fewer spurious reconstructions (684 vs. 8,056). This translates into an approximately 56× higher correct-to-spurious reconstruction ratio, demonstrating VGIA's substantial practical advantage even when exact certification is no longer guaranteed.

Varying client participation

VGIA targets a specific client and does not require its participation in consecutive communication rounds. Intermittent participation therefore does not affect either our theoretical guarantees or experimental results, provided that the attack budget of TT rounds counts only those rounds in which the targeted client participates. In particular, under FedSGD, the guarantees of Proposition 4.1 remain valid as long as its assumptions are satisfied.

We also note that, under our malicious-server threat model, the server controls client selection and can therefore repeatedly select the targeted client whenever it is available.

Varying batches

Our FedAvg experiments already use shuffled minibatches, so batches change across both local steps and communication rounds. As in the original FedAvg formulation (McMahan et al., 2017), each client performs at least one complete local epoch, ensuring that all its records contribute to each update. If the reviewer instead refers to clients processing only a subset of their local dataset in each communication round, this introduces additional challenges. We believe that VGIA's certification guarantees could be extended to this setting and would be happy to detail the proposed procedure if this is indeed what the reviewer intended. However, this extension is not currently covered in the paper and would entail two complications: (i) a potentially longer search phase, since not all records contribute to every update, and (ii) the risk of incorrectly certifying empty slices when the records they contain do not contribute to consecutive updates.

AuthorsOpenReview ↗

Answer to Weakness 3 and Question 3

We thank the reviewer for suggesting a matched end-to-end experiment jointly evaluating reconstruction accuracy, false-certification rates, and detectability under the same neuron and communication budgets.

Appendix G.4 already investigates reconstruction with restricted attack subnetworks (Table 1) and detectability (Tables 2–5), under conditions particularly favorable to the defender. Indeed, while Carletti et al. (2026) propose predefined detection thresholds without a client-specific calibration procedure, we strengthen their methodology by constructing empirical benign distributions from 100 independently trained models on the client's own dataset, calibrating the statistical tests at a nominal 5% false-positive rate. This gives the defender extensive client-specific reference information that would be costly to obtain in realistic FL deployments. Moreover, our controlled setting excludes additional variability arising from heterogeneous client datasets and global aggregation, which could further complicate detection. Nevertheless, we agree that jointly evaluating reconstruction and detectability under matched attack configurations would strengthen our analysis.

We therefore conducted additional experiments on CIFAR-10 (B=2048B=2048) using double precision, jointly evaluating reconstruction and detectability with Na∈{10,50,100}N_a\in\{10,50,100\} attacked neurons per round and direction perturbation p=10−7p=10^{-7} (Appendix G.3). Detection thresholds are calibrated at a nominal 5% false-positive rate using 100 benign (i.e., without any attack) local training simulations. Each configuration is evaluated over 100 seeds, with a maximum of 100, 200, and 1,000 communication rounds for Na=100,50,10N_a=100,50,10, respectively.

The table below reports the average number of communication rounds until VGIA terminates (i.e., produces no further reconstructions), the recovery and false-certification rates, and the false-negative rates of the different hypothesis tests, defined as the proportion of trials in which each test fails to detect the attack. The results reveal a clear trade-off between attack duration and detectability. With Na=100N_a=100, VGIA recovers 99.68% of the samples in approximately 93 rounds, consistently with Table 1, but remains detectable through neuron diversity and entropy. Reducing NaN_a to 50 preserves near-complete recovery (99.71%) in approximately 178 rounds, while substantially improving entropy-based evasion. With Na=10N_a=10, all evaluated detectors exhibit false-negative rates above 95%, while VGIA still recovers 99.75% of the samples in approximately 951 rounds. The false-certification rate remains below 2.25% in all configurations. Notably, the product of the number of attacked neurons per round and the attack duration remains approximately constant across configurations, suggesting that the total number of queried hyperplanes is the key determinant of reconstruction efficiency and that additional communication rounds can effectively compensate for fewer attacked neurons per round.

Overall, these experiments demonstrate that, by restricting the attacked subnetwork, VGIA can achieve near-complete recovery with negligible false-certification rates while substantially reducing its detectability, at the cost of additional communication rounds, even under the defender-favorable calibration assumptions described above.

Matched reconstruction and detectability on CIFAR-10 (B=2048B=2048), for float64 precision. Rounds used and recovery rate report mean ± SD; False certification rate denotes incorrectly isolated samples among all reconstructions. Detection columns report FNR (%) at a nominal 5% FPR, with cumulative cosine similarity evaluated at the end of the attack. Results are averaged over 100 attack seeds.

NaN_aRounds usedRecovery rate (%)False certification (%)Diversity FNR (%)Entropy FNR (%) 10 / 100 binsCosine FNR (%)Cumulative cosine FNR (%)
10092.98 ± 6.1199.68 ± 0.152.227 ± 3.2980.0047.95 / 56.7393.4493.00
50178.44 ± 7.7099.71 ± 0.160.864 ± 1.87342.0196.81 / 97.4392.2988.00
10951.23 ± 12.9099.75 ± 0.130.096 ± 0.08396.3796.30 / 96.1296.1596.00

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.