acceptodds
Withdrawn from ICLR 2027

Protect What Matters, Hide Where: Selective Differential Privacy with Encrypted Masks for Federated Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated Learning (FL) enables collaborative model training without sharing raw data, yet exchanging model updates still poses privacy leakage risks. Existing privacy-preserving methods based on Homomorphic Encryption (HE) or Differential Privacy (DP) often entail substantial computational and communication overhead or result in degraded model utility, respectively. To address these challenges, we propose FedHPP, an efficient hybrid privacy-preserving FL framework that combines model sparsity-aware DP protection with HE-encrypted privacy mask techniques. Generally, FedHPP identifies a block of critical parameters using a Top- selection mechanism based on local model update magnitudes and applies DP noise exclusively to this block. Simultaneously, it uses HE to protect compact binary masks that pinpoint DP-injected and Non DP-injected parameter positions, concealing clients' private parameter selection patterns while substantially reducing encryption overhead. Aggregated masks guide clients in denoising, mitigating the impact of perturbations on model convergence. Extensive experiments conducted on real-world datasets under varying degrees of data heterogeneity and comparison against 10 baselines, demonstrating FedHPP's effectiveness in balancing three-way tensions among FL privacy, utility, and efficiency.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Federated Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_xhtPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes FedHPP, a hybrid privacy-preserving framework for federated learning that combines selective Gaussian noise injection with homomorphically encrypted binary masks. After clipping each client update, FedHPP partitions it into blocks and selects the top-(K) blocks according to an importance score. Gaussian noise is added to the selected blocks, while a binary mask is encrypted using CKKS and aggregated in the ciphertext domain. Clients subsequently decrypt the aggregated mask and use it to gate noise removal. The paper provides theoretical analyses of differential privacy, mask confidentiality, and nonconvex convergence, and evaluates the framework on seven datasets, three model architectures, and ten baselines. The reported results suggest that FedHPP achieves accuracy close to FedAvg while substantially reducing homomorphic-encryption overhead and ciphertext communication compared with selected HE-based baselines. The paper also evaluates its resistance to gradient inversion attacks.

The main question is whether the theoretical guarantees and privacy claims accurately characterize the mechanism implemented by FedHPP. In particular, the relationship between selective noise injection and the claimed differential privacy guarantee, as well as the extent to which encrypted masks conceal the locations of perturbed blocks from an honest-but-curious server, requires further clarification.

Critical strengths
  1. An important problem and a potentially useful systems objective. The paper addresses the practical overhead of combining differential privacy with homomorphic encryption in federated learning. Selectively applying noise and reducing encrypted communication are reasonable directions for improving the efficiency of privacy-preserving training.

  2. Broad empirical evaluation. The evaluation covers multiple datasets, model architectures, and baseline methods, and considers utility, communication overhead, encryption-related computation, and gradient inversion attacks. This breadth provides a useful starting point for evaluating the practical trade-offs of the proposed framework, assuming the privacy guarantees and implementation details can be clarified.

Critical weaknesses
  1. The differential privacy analysis does not appear to establish the claimed guarantee for the implemented selective-noise mechanism. Theorem 1 assumes positive noise variance for every block, whereas the algorithm adds Gaussian noise only to selected top-(K) blocks and releases the remaining coordinates without noise. As written, the theorem therefore appears to analyze a different mechanism from the one deployed. Under standard differential privacy definitions, releasing data-dependent coordinates without noise generally cannot provide the claimed guarantee for the complete update unless additional conditions or protections are established. Moreover, the top-(K) selection itself depends on the client update, and the paper does not clearly establish how the privacy leakage from this selection is accounted for. The noise calibration also requires clarification: whether the stated sensitivity bound is valid depends on the neighboring-dataset definition, the clipping operation, and whether the analysis concerns add/remove or replace-one adjacency. These issues affect the validity of the paper's central privacy claim, rather than merely the tightness of its bound.

  2. The confidentiality of the perturbed-block locations is not established against the stated server adversary. The paper argues that homomorphic encryption protects the binary masks, but the server is also assumed to have access to the clients' uploaded updates. If selected coordinates receive Gaussian noise with a standard deviation of approximately 2.1 under the default parameters, while individual clipped update coordinates are typically much smaller, the resulting magnitude distributions may be readily distinguishable. A simple thresholding attack on the visible updates could therefore reveal which coordinates were perturbed, even without decrypting the masks. The paper does not appear to evaluate this inference channel or explain how the encrypted-mask construction prevents it. Consequently, the cryptographic confidentiality of the mask ciphertext alone is insufficient to establish the broader claim that the locations of perturbed blocks remain hidden.

Questions for authors
  1. Differential privacy of the implemented mechanism. Can the authors provide a formal differential privacy proof for the actual algorithm, in which only the selected top-(K) blocks receive Gaussian noise and the remaining coordinates are released without noise? How is the data-dependent top-(K) selection accounted for in the privacy analysis? Please also specify the neighboring-dataset definition and derive the corresponding sensitivity bound used to calibrate the noise. If the current proof applies only to a full-support Gaussian mechanism, would the authors revise the privacy claim for the implemented selective-noise mechanism?

  2. Inference of perturbed-block locations. Given the stated noise scale and the server's access to uploaded client updates, have the authors evaluated a simple magnitude-thresholding or distribution-based attack to infer which blocks were perturbed? What prevents an honest-but-curious server from distinguishing noisy coordinates from unperturbed coordinates without decrypting the masks? Please report the attack's inference accuracy under the default parameters and discuss its implications for the stated threat model.

  3. Mask-guided denoising and convergence. What is the exact role of the aggregated binary mask in the denoising procedure if each client can reconstruct its own noise from its local random seed? Please clarify the mask aggregation semantics, including whether the mask indicates selection by at least one client or some other condition. How does the method handle the possibility that different clients remove different noise components, and does Theorem 3 account for the resulting update differences and the full denoising procedure?

  4. Importance scoring and comparisons. Why is the signed block sum used as the importance score instead of a magnitude-based statistic that avoids cancellation between positive and negative entries? How sensitive are the results to this design choice? In addition, can the authors clarify the precise technical differences between FedHPP and CLOVER, FedPHE, and FedSel with respect to privacy guarantees, mask confidentiality, and communication and computation overhead?

3 comments

Q1. Privacy of the Actual Release, Adjacency, and Calibration

We thank the reviewer for the insightful and constructive guidance. We clarify and tighten the scope, adjacency definition, and privacy proof framework of Theorem 1 as follows:

  • Adjacency Definition: We adopt the standard record-level Replace-one adjacency definition. Two local client datasets DiD_i and Di′D_i' are adjacent (Di≃Di′D_i \simeq D_i') if they are of the same size and differ in at most one training record.
  • Global Sensitivity Bound: In our single-round analysis, we assume the input global model and other clients' datasets remain fixed. Let u(Di)=ClipC(Δit(Di))u(D_i) = \text{Clip}_C(\Delta_i^t(D_i)) denote the clipped full local update. Due to the clipping operation, ∥u(Di)∥2≤C\|u(D_i)\|_2 \le C holds for any input; thus, the global ℓ2\ell_2 sensitivity upper bound is given by Δ2(u)=sup⁡Di≃Di′∥u(Di)−u(Di′)∥2≤2C\Delta_2(u) = \sup_{D_i\simeq D_i'} \|u(D_i)-u(D_i')\|_2 \le 2C.
  • Selected-Subspace Privacy Guarantee: Although Top-KK sparsity is employed in system engineering implementation, we adopt a conservative sensitivity constraint over the entire update in our formal privacy analysis. Building upon Theorem 1, we derive a more rigorous Selected-Subspace Privacy Guarantee: We first consider the KK selected parameter blocks. Let fS(Di)=[Δˉi,jt(Di)]j∈Sit∈RKBf_{\mathcal{S}}(D_i)=[\bar{\Delta}_{i,j}^t(D_i)]_{j\in\mathcal{S}_i^t}\in\mathbb{R}^{KB} denote the concatenated selected blocks. For a selection set Sit\mathcal{S}_i^t, the corresponding Gaussian mechanism is GS(Di)=fS(Di)+ZS\mathcal{G}_{\mathcal{S}}(D_i)=f_{\mathcal{S}}(D_i)+Z_{\mathcal{S}}, where ZS∼N(0,σ2IKB)Z_{\mathcal{S}}\sim\mathcal{N}(0,\sigma^2 I_{KB}). For a parameter-block selection set Sit\mathcal{S}_i^t, suppose each selected block satisfies ∣Δˉi,jt∣2≤C\\|\bar{\Delta}_{i,j}^t\\|_2\le C. The Gaussian mechanism GS(Di)=fS(Di)+ZS\mathcal{G}_{\mathcal{S}}(D_i)=f_{\mathcal{S}}(D_i)+Z_{\mathcal{S}} satisfies ρK\rho_K-ZCDP [1, 2], where ρK=2KC2σ2\rho_K=\frac{2KC^2}{\sigma^2}. Consequently, for any δ∈(0,1)\delta\in(0,1), it satisfies (ϵK,δ)(\epsilon_K,\delta)-DP with ϵK=ρK+2ρKln⁡(1/δ)\epsilon_K=\rho_K + 2 \sqrt{\rho_K\ln(1/\delta)}.

[1] B. Borja, et al. Improving the gaussian mechanism for differential privacy: Analytical calibration and optimal denoising. ICML, 2018.

[2] B. Mark, et al. Concentrated differential privacy: Simplifications, extensions, and lower bounds. TCC, 2016.

Q2. Inference of Perturbed Data Block Positions

We thank the reviewer for proposing the attack strategy of inferring perturbation positions via magnitude thresholds. We further clarify the noise calibration method, server-observable information, and the evaluation scope of position inference attacks:

  • Noise Standard Deviation Calibration: In the differential privacy mechanism, the noise standard deviation σt\sigma_t is determined by the mechanism sensitivity, privacy parameters, and noise calibration scheme. Specifically, σ\sigma is computed as σ=Δf⋅2ln⁡(1.25/δ)ϵ\sigma = \frac{\Delta f \cdot \sqrt{2\ln(1.25/\delta)}}{\epsilon}, where Δf\Delta f is the ℓ2\ell_2 sensitivity, ϵ\epsilon is the privacy budget, and δ\delta is the failure probability (typically set to 10−510^{-5}).
  • Post-Processing Property of Differential Privacy: Under Non-IID settings, clients' parameter update magnitudes and variances may vary significantly, and updates evolve dynamically across training rounds. The privacy guarantee provided by FedHPP adheres strictly to differential privacy and its post-processing property. For a randomized mechanism M\mathcal{M} satisfying (ϵ,δ)(\epsilon,\delta)-DP, for any adjacent datasets DD and D′D' and any measurable event SS, it holds that Pr⁡[M(D)∈S]≤eϵPr⁡[M(D′)∈S]+δ\Pr[\mathcal{M}(D)\in S] \le e^\epsilon \Pr[\mathcal{M}(D')\in S] + \delta.
  • Threat Model and Threshold Attack Validation: Our threat model accounts for the Level 0 and Level 1 attack capabilities defined in [3]. Under this threat model, since the server lacks access to the exact parameter partitioning details (block size BB) and noisy versus unperturbed blocks are randomly interleaved in the plaintext space, it is difficult for the server to distinguish noisy blocks solely via magnitudes. To evaluate the server's position inference capability, we conducted threshold attack tests based on normalized ℓ2\ell_2 norms (where the server randomly guesses block size BB and computes the accuracy of guessing noisy block positions):
BCIFAR-10(%)CIFAR-100(%)Tiny-ImageNet(%)
5123.782.403.09
10245.664.676.15
20486.028.417.18

The experimental results demonstrate that the server's inference accuracy remains at an extremely low level (around 3% to 8%), thereby strongly validating the confidentiality of the selection patterns.

[3] J. Du, et al. SoK: On Gradient Leakage in Federated Learning. USENIX Security, 2025.

Q3. Mask-Based Denoising and Convergence Analysis

We thank the reviewer for the constructive suggestions regarding aggregated mask semantics, client-side noise reconstruction, and post-denoising convergence analysis.

  • Aggregated Mask Semantics: FedHPP aggregates binary selection masks via block-wise summation: Sjt=∑i∈Itmi,jtS_j^t = \sum_{i\in\mathcal{I}_t} m_{i,j}^t. This variable represents the number of clients adding noise to block jj in round tt. When Sjt=1S_j^t = 1, exactly one client adds noise, enabling client-side local denoising (wˉjt+1−αimi,jt1{Sjt=1}Zi,jt\bar{w}_j^{t+1} - \alpha_i m_{i,j}^t \mathbf{1}\{S_j^t=1\} Z_{i,j}^t). When Sjt>1S_j^t > 1, multiple clients add noise simultaneously, and FedHPP bypasses this correction to preserve aggregated noise statistics.

  • Ablation Study and Convergence: To evaluate the practical gains of the denoising mechanism, we conducted ablation experiments by disabling the denoising operation:

    VariantPartitionCIFAR-10(%)CIFAR-100(%)Tiny-ImageNet(%)
    FedHPPIID74.9336.5734.13
    FedHPP w/o denoisingIID72.1335.2232.48
    FedHPPNon-IID70.9231.0429.87
    FedHPP w/o denoisingNon-IID68.4329.9328.02

    The ablation results fully demonstrate that the mask-guided denoising mechanism consistently enhances model performance across diverse data distributions and secures convergence efficiency under high privacy constraints.

  • Theorem 3 Proof Remarks: In the proof provided in Appendix C.3, we explicitly model the residual aggregated noise term ξrawt=∑i=1Nαi(Mit⊙Zit)\xi_{\text{raw}}^t = \sum_{i=1}^{N} \alpha_i (M_i^t \odot Z_i^t) and its variance bound vt2=ρdσt2∑i=1Nαi2v_t^2 = \rho d \sigma_t^2 \sum_{i=1}^{N} \alpha_i^2. The proof thoroughly accounts for stochastic gradient variance, client heterogeneity, and residual noise effects, rigorously deriving the convergence rate bound of FedHPP.

Q4. Block-Importance Scoring and Comparison with Prior WorkBlock-Importance Scoring:

We adopt the "mean absolute magnitude" (B−1∑b∥xj,b∥B^{-1}\sum_b\|x_{j,b}\|) to prioritize blocks with significant net updates while preventing positive and negative gradient updates from canceling each other out, thereby precisely capturing core optimization directions. We have uniformly revised the corresponding description in the manuscript to "mean absolute magnitude".

Comparison with FedSel, FedPHE, and CLOVER:

we compared FedHPP with methods from other existing works. FedSel achieves local differential privacy through private dimension selection and numerical perturbation, with a privacy analysis that accounts for both the dimension selection process and the protection of the released numerical values. FedPHE employs CKKS-based packed homomorphic encryption (HE) and utilizes secure aggregation and client selection mechanisms to reduce the computational and communication overhead of encryption in heterogeneous federated learning, while also mitigating the impact of straggler clients. CLOVER uses a three-server protocol and secret-sharing techniques to perform secure aggregation while protecting both the values ​​and indices of sparse updates; it further achieves differential privacy goals through distributed noise generation. In contrast, FedHPP does not apply HE to the model parameter values ​​themselves. Instead, it uploads complete, selectively perturbed updates and uses encrypted masks during aggregation to provide selection-count information, which enables local denoising at the client side. The specific differences are as follows:

MethodPrivacy Scope & Positional InfoPrimary Computation & Communication OverheadComplexity (DP+HE)HE Communication Overhead
FedSelAssigns local DP budgets for dimension selection and selected values separately.Involves private selection and value perturbation; communication load depends on sparse encoding.O(d+0)\mathcal{O}(d + 0)O(dbp)\mathcal{O}(d b_p)
FedPHEProtects selected packets using CKKS-based packed HE; the server obtains selection masks.Involves data packing, blinding, encrypted aggregation, and client selection traffic.O(0+pd)\mathcal{O}(0 + p d)O(pdbc)\mathcal{O}(p d b_c)
CLOVERUtilizes three-server secret sharing to protect sparse indices and values; applies DP to aggregated results under specific assumptions.Client traffic scales with sparse workload; secure aggregation and noise generation incur server-side inter-communication overhead.O(d+0)\mathcal{O}(d + 0)O(dbc)\mathcal{O}(d b_c)
FedHPPAdopts dense selective perturbation updates and encrypted block masks; updates may risk positional information leakage.Involves truncation, scoring, noise sampling, and HE-based mask handling. Uploads dd values plus ≈⌈J/s⌉\approx \lceil J/s\rceil ciphertexts (with packing capacity ss).O(d+dB)\mathcal{O}(d + \frac{d}{B})dbp+⌈dB⌉bcd b_p + \lceil \frac{d}{B} \rceil b_c

The authors have addressed some of my concerns. However, it remains unclear to me whether DP can be guaranteed when only top-K coordinates are selected,while the remaining coordinates are released without noise, given that a local model updates typically changes across all coordinates.

Reviewer_qRYhOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper attempts to propose a design of a hybrid solution with homomorphic encryption (CKKS) and differential privacy to reduce the computational cost from CKKS for privacy protection in FL via selective masking.

Critical strengths

The motivation of addressing the expensive cost of full HE protection for FL is clearly stated and has been researched widely for the domain.

Critical weaknesses

My core concern for this paper is lack of novelty and missing proper previous work. Specifically, the core problem formulation and proposed solution has been widely studied in the research community already. For example, FedML-HE [1] from 2023 (missing reference in this submission) already proposed a selective encryption solution via sensitivity analysis to reduce computational and communication overheads which combines the protection of HE and DP. Additionally, their work provides formal quantification of privacy loss from selective encryption, which this work lacks. The novelty of this work is marginal.

[1] Jin, Weizhao, et al. "FedML-HE: An efficient homomorphic-encryption-based privacy-preserving federated learning system." NeurIPS Federated Learning Workshop (2023).

Questions for authors

As pointed out in weaknesses.

2 comments

We thank the reviewer for recommending the important work FedML-HE [1]. We have included citations and discussions regarding this work in the revised manuscript. Although both approaches share similar optimization goals, notable differences remain in their protection targets, protocol workflows, and model update mechanisms:

  1. Different Roles of HE and Perturbation:FedML-HE utilizes parameter sensitivity analysis to select model parameters for encryption, thereby reducing HE overhead through selective parameter encryption, and its formal analysis also discusses the mechanism of applying Laplace perturbation to unencrypted parameters. In contrast, FedHPP does not encrypt model parameter values via HE; instead, it adopts selective Gaussian perturbation on full model updates and employs HE solely to protect block selection masks. This architecture avoids performing HE aggregation on high-dimensional model parameters, but it comes with a clear security boundary: unperturbed parameters remain visible to the server. Consequently, FedHPP does not claim equivalent confidentiality to the encrypted parameters in FedML-HE, nor does it directly equate selective Gaussian perturbation with standard DP applied to full updates.
  2. Mask-Guided Client-Side Denoising:Another distinct feature of FedHPP is the introduction of a client-side local noise cancellation mechanism driven by the aggregated selection masks. Specifically, clients first select Top-KK blocks based on local updates and add noise; the server then aggregates the perturbed updates and encrypted masks separately. Trusted clients utilize aggregated mask counts to identify blocks with unique noise contributors and leverage locally stored random seeds to remove the noise components they individually introduced in those blocks. This design does not attempt to recover other clients' private noise, nor does it yield a fully denoised global model at the server side. Its primary purpose is to mitigate the perturbation impact on clients' local models under specific conditions.
  3. Theoretical Formalization of Privacy Loss:Due to the differences in protection targets and information disclosure modes between FedML-HE and FedHPP, their formal analyses cannot directly substitute for one another. FedHPP's security analysis mainly consists of privacy mechanism analysis tailored for selective noise addition and computational indistinguishability of mask ciphertexts based on the semantic security of the underlying HE scheme.

To further illustrate the differences between the two designs in practical systems, we provide a quantitative comparison of FedHPP and FedML-HE over a single training round (taking CIFAR-10, block size = 1024, Top-KK = 0.5 as an example):

MethodPartitionAccDP-TimeDP-TrafficHE-TimeHE-TrafficPSNR(CFGIA)
FedHPPNon-IID70.920.042.410.0040.326.65
FedML-HENon-IID68.210.0041.212.961686.527.02

Under identical configurations, FedHPP achieves 70.92% accuracy, outperforming FedML-HE (68.21%) by 2.71 percentage points. FedHPP significantly reduces HE overhead (0.004 s / 0.32 MB vs. 2.960 s / 1686.52 MB) by applying HE only to mask metadata, though this incurs slightly higher DP costs due to selective perturbation and full-model updates. This highlights distinct cost-allocation strategies rather than an absolute advantage in all metrics. Under the CFGIA attack, both frameworks demonstrate comparable privacy protection (PSNR: 6.65 dB vs. 7.02 dB).

In addition to theoretical guarantees, we have reported in the revision the PSNR, SSIM, and LPIPS metrics under iDLG, InvG, FedLeak, and CFGIA attacks to validate the defense efficacy of "selective differential privacy." Meanwhile, we have also provided utility results under non-IID data partitions to support our discussion on the privacy–utility trade-off.

In summary, we more accurately characterize the main contribution of FedHPP as: a hybrid privacy-preserving framework combining block-level selective Gaussian perturbation, encrypted selection metadata, and client-side local noise correction. The primary research focus of this paper is not on being the first to apply DP and HE, but rather on exploring the synergistic relationships among different protection operations and improving the practical trade-off among privacy defense, model utility, and system efficiency under data heterogeneity scenarios.

[1] W. Jin, et al. FedML-HE: An efficient homomorphic-encryption-based privacy-preserving federated learning system. NeurIPS Federated Learning Workshop, 2023.

I thank the authors for their rebuttal and the additional clarifications. However, I believe my main concerns about meaningful research contribution have not been fully addressed, and I will therefore maintain my original rating.

Reviewer_QP7EOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FedHPP adds Gaussian noise to selected parameter blocks and homomorphically encrypts masks identifying those blocks. Clients use aggregated masks to remove their own noise in certain uniquely perturbed blocks. Encrypting compact masks is a reasonable efficiency objective. However, the central differential privacy theorem assumes a different mechanism from the proposed algorithm: it requires full-support, data-independent noise, whereas FedHPP selects noise locations from private updates and releases unselected coordinates without perturbation. This undermines the claimed privacy–utility comparison.

Critical strengths
  1. The paper addresses a relevant systems objective, since encrypting compact metadata instead of entire model updates can substantially reduce cryptographic computation and transmitted ciphertext.
  2. The experiments cover multiple datasets, model architectures, heterogeneity settings, and reconstruction attacks.
Critical weaknesses
  1. Theorem 1 does not establish differential privacy for the proposed algorithm. Section 4.2, Equation 2, and Algorithm 1 perturb only Top-K blocks, whereas Theorem 1 and Appendix C.1 require positive variance in every block and a fixed covariance independent of private data. For example, if neighboring datasets produce clipped two-block updates of (0.5, 0.1) and (0.5, 0.2), with K = 1, both select the first block. The second uploaded coordinate is then exactly 0.1 or 0.2. The event that it equals 0.1 has probability one under one dataset and zero under the other, ruling out the stated approximate DP guarantee for delta below one unless additional assumptions exclude this case. Encrypting the mask does not conceal that coordinate.
  2. Ciphertext security alone does not establish the claimed mask confidentiality. The server observes selectively perturbed updates as well as encrypted masks, and statistical differences between noisy and unperturbed blocks may reveal selection patterns. Semantic security of the ciphertext does not rule out inference from the server’s complete observation. Theorem 2 should use an appropriate indistinguishability formulation rather than claim that a potentially predictable mask cannot be recovered.
  3. The contribution of mask-guided denoising is not demonstrated. Equation 4 removes a client’s own noise only when that client uniquely selected the block; other clients and the current server model retain that noise. The empirical frequency of this condition should be reported. Ablations disabling denoising and mask transmission are needed to establish their benefit.
  4. The experiments do not demonstrate an advantage under matched privacy guarantees. If the claimed DP guarantee is invalid, higher accuracy at the same nominal epsilon cannot establish a better DP privacy–utility trade-off. Reconstruction failures under selected attacks cannot replace a formal guarantee. Attack budgets, restarts, attacker knowledge, and attacks specifically targeting unperturbed coordinates should be disclosed and evaluated.
  5. Additional specifications need correction. The block importance formula uses a signed mean despite being described as an average magnitude, permitting positive and negative updates to cancel. The convergence bound also assumes, rather than derives, a bound on the effective update direction error, limiting the conclusions that can be drawn about the benefits of the proposed design.
Questions for authors

Does the implementation add nonzero noise to unselected blocks? If not, how is the counterexample above excluded? What mechanism does Theorem 1 certify, and does that mechanism match the reported experiments? How is the total privacy budget computed across all 200 rounds and observable releases? What mask-confidentiality property is claimed when the server observes both ciphertexts and selectively perturbed updates? How frequently does the unique-contributor denoising condition occur, and what do the no-denoising and no-mask ablations show? Is block importance computed using a signed mean, a mean absolute value, or another statistic?

3 comments

Q1. Noise on Unselected Blocks and the Counterexample.

Thank you for your valuable comments. To clarify, FedHPP does not upload selected and unselected blocks as independent messages separately. Specifically:

  • Clients first apply gradient clipping to the full local model update to bound its overall magnitude, and then apply Gaussian perturbation to the Top-KK parameter regions selected based on block importance. Unselected parameter blocks receive no additional noise, but all parameters are ultimately uploaded as a complete model update vector, with no explicit block selection tags or positional identifiers included in the communication format.
  • The reviewer's two-block counterexample illustrates that if a data-dependent coordinate is never perturbed, its precise publication may violate standard DP definitions. However, this counterexample abstracts away a simplified coordinate scenario and fails to capture the complex parameter statistical distributions, Non-IID data, and multi-round stochastic optimization processes inherent in high-dimensional FL. Therefore, one cannot directly deduce that the server can accurately identify the size and range of perturbed regions in real training environments, let alone successfully execute DLG. We have included additional comparisons in Table 3 regarding the defense capabilities of other DPFL schemes against reconstruction attacks, validating our defense efficacy.

Q2. Scope of Theorem 1

Theorem 1 establishes the privacy mechanism for selectively perturbed parameter blocks. Although Top-KK sparsity is employed in system engineering implementation, we adopt a conservative sensitivity constraint over the entire update in our formal privacy analysis. Building upon Theorem 1, we derive a more rigorous Selected-Subspace Privacy Guarantee: We first consider the KK selected parameter blocks. Let fS(Di)=[Δˉi,jt(Di)]j∈Sit∈RKBf_{\mathcal{S}}(D_i)=[\bar{\Delta}_{i,j}^t(D_i)]_{j\in\mathcal{S}_i^t}\in\mathbb{R}^{KB} denote the concatenated selected blocks. For a selection set Sit\mathcal{S}_i^t, the corresponding Gaussian mechanism is GS(Di)=fS(Di)+ZS\mathcal{G}_{\mathcal{S}}(D_i)=f_{\mathcal{S}}(D_i)+Z_{\mathcal{S}}, where ZS∼N(0,σ2IKB)Z_{\mathcal{S}}\sim\mathcal{N}(0,\sigma^2 I_{KB}). For a parameter-block selection set Sit\mathcal{S}_i^t, suppose each selected block satisfies ∣Δˉi,jt∣2≤C\\|\bar{\Delta}_{i,j}^t\\|_2\le C. The Gaussian mechanism GS(Di)=fS(Di)+ZS\mathcal{G}_{\mathcal{S}}(D_i)=f_{\mathcal{S}}(D_i)+Z_{\mathcal{S}} satisfies ρK\rho_K-ZCDP [1, 2], where ρK=2KC2σ2\rho_K=\frac{2KC^2}{\sigma^2}. Consequently, for any δ∈(0,1)\delta\in(0,1), it satisfies (ϵK,δ)(\epsilon_K,\delta)-DP with ϵK=ρK+2ρKln⁡(1/δ)\epsilon_K=\rho_K + 2 \sqrt{\rho_K\ln(1/\delta)}.

[1] B. Borja, et al. Improving the gaussian mechanism for differential privacy: Analytical calibration and optimal denoising. ICML, 2018.

[2] B. Mark, et al. Concentrated differential privacy: Simplifications, extensions, and lower bounds. TCC, 2016.

Q3. Mask Confidentiality under Joint Observations

Thank you for your feedback; we would like to provide the following clarifications.

  • Mask-Ciphertext Confidentiality: In FedHPP, clients protect binary block selection masks using HE. Under the conditions that the underlying HE scheme satisfies IND-CPA security, the server has no access to the decryption private key, and the key management protocol is correctly executed, the encrypted masks achieve computational indistinguishability.
  • Position Inference Resistance: The threat model for FedHPP considers the Level 0 and Level 1 attack capabilities attributed to the server in [3]. Under uncrackable mask ciphertexts, it is difficult for the server to deduce specific noisy regions simply by observing uploaded model parameters. Due to the inherent heterogeneity of Non-IID data and the effects of dynamic gradient clipping, noisy and unperturbed blocks are not arranged in an orderly fashion in the plaintext space, but rather form dynamic random sequences. Furthermore, without knowing the exact parameter partitioning (block size BB), the server cannot accurately infer block boundaries.
  • Experimental Validation: To evaluate the server's capability in positional mask inference, we conducted threshold attack tests (where the server randomly guesses block size BB and computes the guessing accuracy of noisy block positions based on normalized ℓ2\ell_2 norms):
BCIFAR-10(%)CIFAR-100(%)Tiny-ImageNet(%)
5123.782.403.09
10245.664.676.15
20486.028.417.18

The results show that simple threshold attacks cannot achieve effective classification, demonstrating that FedHPP successfully preserves the concealment of selection patterns via its multi-dimensional confusion mechanism. Accordingly, we have refined Theorem 2 in the revised manuscript to strictly restrict the formal security conclusion to the computational indistinguishability of mask ciphertexts provided by the underlying HE.

[3] J. Du, et al. SoK: On Gradient Leakage in Federated Learning. USENIX Security, 2025.

Q4. Unique Contributors and Ablation of Denoising & Mask Transmission

  • Definition and Statistics of Unique Contributors: FedHPP first computes the sum of the binary masks from all clients for each block, yielding an aggregated mask count: Sjt=∑i∈Itmi,jtS_j^t = \sum_{i\in\mathcal I_t} m_{i,j}^t, where It\mathcal I_t denotes the set of clients participating in training during round tt. When Sjt=0S_j^t = 0, no client has added noise to block jj; when Sjt=1S_j^t = 1, exactly one client has added noise to that block; and when Sjt>1S_j^t > 1, multiple clients have added noise to the same block. To determine the actual frequency with which the "unique contributor" condition occurs, we employ the following statistical definition: ∑t∑i∈It∑jmi,jt1Sjt=1∑t∑i∈It∑jmi,jt\frac{\sum_t\sum_{i\in\mathcal I_t}\sum_j m_{i,j}^t \mathbf{1}{S_j^t=1}}{\sum_t \sum_{i\in\mathcal I_t}\sum_j m_{i,j}^t}. This metric represents the proportion of all noise-added blocks that satisfy the unique contributor condition, thereby enabling local denoising. Measured frequencies for representative clients are as follows:

    ID1050100150200
    Client 10.38%0.57%1.13%0.57%2.08%
    Client 20.25%0.95%0.38%0.76%1.32%
  • Denoising Mechanism Ablation Study:

    VariantPartitionCIFAR-10(%)CIFAR-100(%)Tiny-ImageNet(%)
    FedHPPIID74.9336.5734.13
    FedHPP w/o denoisingIID72.1335.2232.48
    FedHPPNon-IID70.9231.0429.87
    FedHPP w/o denoisingNon-IID68.4329.9328.02

    The ablation results fully demonstrate that the mask-guided denoising mechanism consistently improves model performance across various data distributions.

  • Regarding Mask Transmission: The encrypted masks serve to support privacy-preserving block selection counting and provide the basis for trusted clients to evaluate denoising conditions. If mask transmission were completely disabled without alternative channels, clients would be unable to perform unique contributor denoising (its performance thus aligns with FedHPP w/o denoising). As shown in Table 2 (e.g., CIFAR-10), mask transmission incurs negligible overhead (HE time of 0.004s and traffic of 0.32 MB).

Q5. Equivalent DP Guarantee and Privacy Budget

In FedHPP's experimental setup, each participating client adopts a nominal per-round privacy budget of ϵt=2\epsilon_{t} = 2 (rather than treating ϵ=2\epsilon = 2 as the cumulative budget across all 200 training rounds).

  • Budget Composition Analysis: Privacy accounting across multiple rounds is tracked using basic sequential composition theorems or the compositional properties of zCDP. If a client participates in TiT_i rounds of training, the basic composition theorem yields the upper bound: ϵitotal≤∑t∈Tiϵt=200×2=400\epsilon_{i}^{\mathrm{total}} \le \sum_{t\in \mathcal{T}_i} \epsilon_{t} = 200 \times 2 = 400, and δitotal≤∑t∈Tiδt\delta_{i}^{\mathrm{total}} \le \sum_{t\in T_{i}} \delta_{t} (finer bounds can also be derived via zCDP).
  • Threat Model and Attack Evaluation: We consider the Level 0 and Level 1 attack capabilities of servers as defined in [1]. We conduct reconstruction attack evaluations (DLG, InvG, and FedLeak) every 10 rounds over the 200 training rounds, measuring quality with PSNR, SSIM, and LPIPS. Note that "evaluating every 10 rounds" represents the experimental evaluation frequency rather than suggesting the server only observes updates during those rounds. Furthermore, launching targeted attacks on unperturbed coordinates is impractical because it would require the malicious server to possess prior knowledge of specific dynamic noise-addition details, thereby violating standard black-box/threat model assumptions.

Q6. Typo Correction

We thank the reviewer for pointing this out. In Section 4.2, we actually employ the "mean absolute magnitude" (B−1∑b∥xj,b∥B^{-1}\sum_b\|x_{j,b}\|); this is intended to prevent positive and negative gradient updates from canceling each other out and to precisely capture core optimization directions. We have uniformly corrected the relevant descriptions in the revised manuscript to "mean absolute magnitude".

I appreciate the authors’ efforts. The response partially addresses my concerns, but I think the core problem remains unresolved. Claims about high-dimensional complexity, Non-IID, and multi-round stochastic optimization are not reasons to dismiss the DP issue. The supplementary attack experiments can only show empirical defense and cannot prove that the algorithm as a whole satisfies DP. Therefore, I will keep my score.

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

Related papers

Discussion (0)

Sign in to comment.