acceptodds
Under review as a conference paper at ICLR 2027

HPFL: Hybrid Privacy-Preserving Federated Learning Framework Based on Homomorphic Encryption and Differential Privacy

Abstract

Federated Learning (FL) enables model training on distributed data without the disclosure of private information. However, it is still vulnerable to malicious aggregators in FL. Existing Differential Privacy (DP) methods suffer from accuracy degradation due to noise injections. While Homomorphic Encryption (HE) schemes have no model accuracy loss, they need tremendous computation overheads, especially for a large model, which slows down FL convergence time. Moreover, existing hybrid DP and HE schemes suffers from low accuracy and high computation complexity. To this end, this paper proposes a hybrid privacy-preserving FL (HPFL) framework that dynamically integrates HE with DP at each device's side. We formulate this issue as a utility maximization problem under the consideration of time consumption for both model training and model protections within each global training round. A closed-form optimal HE ratio for each device is theoretically derived, and a model gradient magnitude-based mask is further designed to determine the HE-based model parameters. Extensive experiments on FMNIST, CIFAR-10, and STL-10 demonstrate that the proposed HPFL significantly outperforms DP method only with accuracy improvements of up to 30%. Moreover, our proposed HPFL can significantly reduce the overall training time of FL by approximately 80%, while suffering from 3% accuracy loss, compared with HE only.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Differential Privacy, Homomorphic Encryption

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

This paper proposes HPFL, a hybrid privacy-preserving federated learning framework that combines homomorphic encryption (HE) and differential privacy (DP). The method dynamically adjusts the HE ratio based on each client's computational and communication resources, and uses gradient magnitudes to select which model parameters should be encrypted. The authors derive an optimal HE ratio and evaluate the method on several datasets, including FMNIST, CIFAR-10, and STL-10.

Critical strengths

I think the paper addresses a practical and interesting problem, as balancing privacy, model accuracy, and computational efficiency is important in federated learning. The idea of dynamically adjusting the HE ratio based on client capabilities is reasonable, especially when clients have different computational resources. I appreciate that the authors include several datasets, ablation studies, and privacy attack evaluations to support their method.

Critical weaknesses
  1. I have some concerns about the hybrid aggregation in Eq. (6). Since different clients may use different masks, it is unclear whether combining the aggregated HE and DP parameters using each client's own mask can correctly recover the global model. Some contributions from other clients seem to be missing.

  2. I am not fully convinced by the privacy analysis in Appendix C. In particular, Eq. (C.1) seems to replace coordinate-wise masking with a scalar DP ratio, which is not generally equivalent. Also, the masks are selected based on local gradients, but I could not find a clear analysis of how this affects the overall privacy guarantee across multiple training rounds.

  3. I find the derivation of the optimal HE ratio somewhat questionable. In Eq. (9), both terms appear to increase with the round duration, so I do not quite see why the objective should have the unique optimum claimed by the authors. I think this part needs a more careful explanation.

  4. There also seems to be an issue with the mask selection in Eq. (12). If the HE ratio is used directly as the quantile threshold, selecting parameters above that threshold would generally give a fraction of approximately 1−hk,pt,1-h_{k,p}^{t,} rather than hk,pt,h_{k,p}^{t,}. Unless the authors are using a different quantile definition, the selected HE ratio may not match the intended one.

Questions for authors
  1. Could the authors explain how the aggregation works when clients use different masks? In particular, how does the method ensure that all clients obtain the same correctly aggregated global model?

  2. Could the authors clarify the privacy guarantee for the complete framework, especially considering the gradient-based mask selection and multiple training rounds?

  3. Could the authors provide more details on why Eq. (9) has a unique optimum? From the current formulation, the objective seems to be monotonically increasing.

  4. Could the authors clarify the quantile definition used in Eq. (12)? How does the resulting mask ensure that the fraction of HE-protected parameters matches the optimal HE ratio?

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

The paper proposes HPFL, a federated learning framework that dynamically partitions model parameters into homomorphically encrypted and differentially private components. It formulates an optimization problem to determine device-specific HE ratios under computation and communication constraints and introduces a gradient-magnitude-based mask to select parameters for encryption. The framework aggregates encrypted and perturbed parameters separately and combines them to construct updated global models. Experiments compare HPFL against privacy-preserving FL baselines and evaluate training accuracy, computation time, component ablations, and resistance to membership inference attacks.

Critical strengths
  1. The paper focuses on a relevant and practically important problem: jointly balancing privacy, model utility, and heterogeneous computation/communication costs in cross-device FL.
  2. The idea of adapting the HE/DP split to the resource condition of each client is well motivated.
  3. The experimental design includes useful ablations for the adaptive HE ratio, parameter importance based mask, and dynamic mask update.
Critical weaknesses
  1. The novelty over existing hybrid HE/DP federated learning approaches is insufficiently established. Combining HE and DP in FL is not new. (Q1-3)
  2. The proposed hybrid aggregation mechanism does not guarantee a consistent global model across devices. (Q4)
  3. The DP guarantee is insufficiently justified. Appendix C uses a scalar protection ratio in its sensitivity derivation, although the actual mechanism uses a data-dependent coordinate mask. Privacy composition over rounds is missing, and the reported ϵ=70\epsilon=70 lacks a corresponding δ\delta, noise multiplier, and final privacy bound. (Q5)
  4. The privacy-budget evaluation is too narrow. Most accuracy results use ϵ=70\epsilon=70, while the appendix evaluates membership inference attacks at ϵ=50\epsilon=50 and 7070. Results under stronger and intermediate privacy settings, with the associated δ\delta, noise calibration, and composed privacy accounting, would show whether HPFL's accuracy and efficiency advantages hold across meaningful privacy requirements. (Q6-7)
  5. The manuscript needs substantial proofreading and editorial revision before publication. Several grammatical and formatting errors include but are not limited to:
  • on page 1, "existing hybrid DP and HE schemes suffers" -> suffer
  • on page 1, "experiments on diverse data setting" -> settings
  • on page 1, "demonstrate that proposed HPFL" -> the proposed
  • on page 1, "we need the design to be a low computation complexity" -> The design must have low computational complexity.
  • on page 2, "others may prefer to DP" -> prefer DP
  • on page 2, "In follows, we discuss…" -> In the following,
  • on page 3, "aims to use DP or HE to different devices" -> apply DP or HE to
  • on page 3, "with s numbers of local iterations" -> with s local iterations
  • on page 7, "at t-th global round" -> at the t-th
  • on page 9, "Table 1 presents the test accuracies of eight FL methods on four datasets", there are only three datasets in Table 1.
Questions for authors
  1. What are the fundamental methodological differences between HPFL and prior selective HE/DP approaches listed below? Beyond combining established privacy mechanisms, what new capabilities does HPFL provide?
  2. What advantages does the proposed resource-aware HE-ratio optimization offer over existing adaptive protection strategies or simpler resource-based allocation heuristics? Could the authors provide controlled comparisons under matched privacy, computation, and communication constraints?
  3. Given that gradient- or Fisher-based parameter importance has already been explored in MaskCrypt and SelectiveShield, what is novel about HPFL's EMA-based parameter selection? Could the authors compare it with representative existing importance metrics and isolate its independent contribution to privacy, accuracy, and efficiency?
  4. How does HPFL guarantee that all participating devices reconstruct the same global model when their binary masks differ? In addition, how are the HE-protected model parameters and DP-protected model updates combined consistently? Could the authors provide a rigorous aggregation formulation and a simple example to clarify how client-specific HE masks are represented, how CKKS ciphertext slots are aligned across clients with different masks, and whether the adopted packing strategy supports the computation and communication cost assumptions in Eqs. (2)–(3)?
  5. Please clarify the exact DP mechanism and its scope. What is the adjacency definition, how is the sensitivity derived for a data-dependent coordinate mask, what are the clipping and Gaussian noise parameters, and what final (ϵ,δ)(\epsilon,\delta) guarantee follows after composition across rounds?
  6. Could the authors evaluate a broader range of privacy budgets, including settings substantially stronger than ϵ=50\epsilon=50 ? For each setting, please report test accuracy, privacy-attack performance, realized HE ratio, and the noise multiplier, δ\delta, and final composed privacy budget so that the privacy-utility trade-off can be compared fairly.
  7. Why do several methods achieve membership inference attack success rates substantially below 50% in Fig. 4(a)? Since reversing binary predictions could increase these accuracies, how should these results be interpreted as evidence of privacy protection? Could the authors report attack performance after accounting for possible label inversion, clarify the attack training and evaluation protocols, and provide consistent AUC-based comparisons against random guessing and an unprotected baseline?

2022, IEEE TII - Jia et al., Blockchain-Enabled Federated Learning Data Protection Aggregation Scheme With Differential Privacy and Homomorphic Encryption in IIoT.

2025, IEEE TDSC - Hu and Li, MaskCrypt: Federated Learning With Selective Homomorphic Encryption.

2025, arXiv - Korkmaz and Rao, A Selective Homomorphic Encryption Approach for Faster Privacy-Preserving Federated Learning.

2025, arXiv - Li et al., SelectiveShield: Lightweight Hybrid Defense Against Gradient Leakage in Federated Learning.

2026, arXiv - Gül et al., SecureDrive-FL: Joint Differential Privacy and Gradient-Aware Selective Homomorphic Encryption for Federated Driver Monitoring.

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

This paper proposes HPFL, a hybrid privacy-preserving federated learning framework that combines Homomorphic Encryption (HE) and Differential Privacy (DP) to balance model accuracy and computational overhead. The authors formulate an accuracy maximization problem under training time constraints and derive a closed-form solution for the optimal HE ratio for each device. A gradient magnitude-based masking mechanism is further designed to selectively protect important model parameters using HE while applying DP to the remaining parameters. Experimental results on FMNIST, CIFAR-10, and STL-10 demonstrate that HPFL achieves higher accuracy than DP-based methods and significantly reduces training time compared to HE-only approaches, with limited accuracy degradation.

Critical strengths

S1. The problem is straightforward and easy to understand, yet it has remained a longstanding challenge.

S2. The authors formulate the HE ratio selection as an optimization problem under training latency constraints and derive a closed-form solution. This provides a principled approach to dynamically allocating privacy protection mechanisms according to device capabilities.

S3. The proposed gradient magnitude-based masking mechanism identifies important model parameters for HE protection. The dynamic mask design is intuitive and is supported by ablation studies comparing it against random and fixed masking strategies.

Critical weaknesses

W1. Optimizing under a fixed total training-time budget does not adequately reflect realistic federated learning scenarios. Moreover, the proposed formulation fails to fully capture device heterogeneity. In practice, heterogeneous computational capabilities across clients lead to different local training times, and the duration of each global training round is typically determined by the slowest participating client (i.e., the straggler). Therefore, the practical relevance and applicability of the proposed optimization framework remain questionable.

W2. The theoretical analysis supporting the claimed optimal HE ratio contains significant mathematical issues. In Section 5.1, the authors claim that the transformed objective in Eq. (9) is unimodal and admits a unique global maximum. However, both terms of the objective appear to be non-decreasing with respect to the round duration, which does not support the existence of a unique finite optimum. Moreover, the capped linear HE-ratio function is incorrectly characterized as convex.

W3. The authors compute the threshold using the quantile corresponding to the HE ratio and select parameters whose importance scores exceed this threshold. Under the standard definition of quantiles, this would select approximately (1-h) rather than (h) of the parameters.

W4. The experiments rely on relatively limited model architectures and largely synthetic device and communication configurations. More comprehensive comparisons with existing selective-encryption approaches, together with statistical variability and realistic end-to-end latency measurements, are needed.

W5. The central problem addressed in this paper is relatively straightforward and has been extensively studied in privacy-preserving federated learning. The proposed framework essentially combines two established privacy mechanisms, HE and DP, and applies gradient-based parameter selection to balance accuracy and computational overhead. Similar ideas involving selective encryption and hybrid privacy protection have already appeared in prior work.

Questions for authors

NA

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.