Pith. sign in

REVIEW 5 major objections 5 minor 40 references

A Privacy-Centric Approach: Scalable and Secure Federated Learning Enabled by Hybrid Homomorphic Encryption

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read FLHHE combines symmetric encryption with homomorphic encryption so federated-learning clients upload 16x less per round than full homomorphic encryption, at the same model accuracy, by moving the transciphering cost to the server.

desk verdict A useful and believable but narrowly-scoped benchmark of HHE for FL; the client-side savings are real, but the scalability claims outrun a one-round, three-client simulation. read the letter →

arxiv 2507.14853 v2 pith:F4OE2FWY submitted 2025-07-20 cs.CR cs.LG

Khoa Nguyen , Tanveer Khan , Hossein Abdinasibfar , Antonis Michalas This is my paper · ORCID
classification cs.CRcs.LG
keywords federatedlearninghybridhomomorphicencryptiontranscipheringRubatocommunicationoverheadprivacy-preservingmachinesecureaggregationCKKS
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes FLHHE, a federated learning protocol that wraps client model updates in a lightweight symmetric cipher (Rubato) and lets the aggregation server homomorphically convert them (transcipher) into ciphertexts it can average without decryption. The central claim is that this hybrid arrangement preserves the exact accuracy of plaintext and full-homomorphic aggregation while cutting each client's upload from 16.8 MB to 1.048 MB per round and trimming client encryption time from 0.21 s to 0.008 s, at the price of about 6.75 s of extra server work per client per round. If the claim holds, FL deployments in bandwidth- or battery-limited settings can obtain encrypted aggregation at a fraction of the client-side cost of pure homomorphic encryption.

What carries the argument

The Rubato/RtF transciphering framework: a stream cipher (Rubato) generates a keystream from a secret key and a nonce; the client adds the keystream to scaled model weights, and the server, holding an FV-encrypted copy of the key, homomorphically reconstructs the keystream, subtracts it, and applies the HalfBoot procedure to convert the result into a CKKS ciphertext that can be averaged. This machinery is what moves the expensive part of encryption off the client while keeping the server's aggregation blind.

What would settle it

Measure FLHHE with more than 3 clients and multiple rounds on distributed hardware: if server-side transciphering time per client does not decrease with parallelization, or if the 16x upload saving is outweighed by the unchanged 16.8 MB download, the practicality claim fails. Also, if running the protocol to convergence on a real-valued task (not one round on MNIST) yields accuracies that diverge from plaintext aggregation, the 'no accuracy degradation' claim would be refuted.

Watch

Extended reading notes

Core claim

Under the honest-but-curious threat model, clients encrypt their trained weights with a symmetric key and send the result along with a nonce; the server uses a homomorphically encrypted copy of the symmetric key to evaluate the Rubato keystream, subtract it from the received ciphertext, and apply a HalfBoot to obtain CKKS ciphertexts carrying the plaintext messages. Averaging these ciphertexts yields the same global model as plaintext averaging: the paper reports identical accuracies of 65.92% overall, 73.15%, 54.35%, and 58.43% on three held-out label subsets for plaintext, HE, and HHE. The trade-off is that server-side transciphering dominates runtime at 6.75 s per client per round, versus 0.004 s for direct HE aggregation and 0.00005 s for plaintext.

Load-bearing premise

The reported savings are measured on a single simulated environment (one laptop, three clients, a single small neural network, one dataset, and apparently one aggregation round), and that environment is taken to be representative of real, larger-scale federated deployments.

Editorial extensions

If this is right

  • Clients upload 16x less per round than full HE (1.048 MB vs 16.8 MB), making encrypted aggregation viable on constrained connections.
  • Client compute overhead falls to 0.008 s for encryption, roughly 26x faster than HE encryption, so local training dominates client runtime.
  • Aggregated model accuracy under HHE exactly matches plaintext and HE, so the privacy gain costs no utility in this setting.
  • Server transciphering at ~6.75 s per client per round becomes the bottleneck, but it is parallelizable and can be precomputed in part (offline phase).

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The 16x savings applies to uploads only; total per-client communication is halved (17.848 MB vs 33.6 MB) because the server still sends back a 16.8 MB HE-encrypted aggregated model, so the download side remains an open bottleneck.
  • Accuracy is demonstrated on a single round with three non-overlapping MNIST partitions; the protocol's behavior across multiple aggregation rounds and heterogeneous data distributions is not measured, so the identity of accuracies is a single-round observation rather than a convergence result.
  • If server-side transciphering were parallelized across many cores or machines, the per-client 6.75 s could drop enough to make HHE-based FL practical at scale; a testable extension would measure throughput on a multi-server cluster.
  • For real-valued models, HHE relies on quantization and the Rubato cipher's 16-bit precision; scenarios needing higher-precision gradients (e.g., some medical or financial models) may require different symmetric ciphers.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. This paper describes FLHHE, a protocol that integrates hybrid homomorphic encryption (HHE) into federated learning. Clients train local models in plaintext, encrypt their model updates with a symmetric cipher, and send the ciphertexts to a server; the server transciphers them into CKKS/HE ciphertexts using the Rubato/RtF framework, performs averaging in the encrypted domain, and returns the HE-encrypted global model. The experimental section compares plaintext, CKKS-based HE, and HHE settings on MNIST with three clients and a two-layer MLP, reporting client upload sizes, client/server computation times, and accuracies. The main claimed results are that HHE reduces client upload by about 16x relative to HE (1.048 MB vs 16.8 MB), adds negligible client encryption time (0.008 s vs 0.21 s), and yields identical accuracies, at the cost of about 6.75 s of server-side transciphering per client per round.

Significance. The paper addresses a real deployment bottleneck in privacy-preserving FL: client-side encryption overhead. If the reported measurements are representative, HHE is a structurally sound way to shift cost from clients to the server while preserving exact aggregation accuracy. The authors should be credited for direct measurements rather than analytic guesswork, for comparing plaintext, HE, and HHE in a common testbed, for reporting the server-side cost honestly rather than hiding it, and for releasing source code. The main limitation is scope: a single small model, three clients, one dataset, no stated training hyperparameters, and no multi-round results. As a case study of one FL aggregation round, the work is informative; as a demonstration of 'scalable and secure' FL, it requires the additional experiments and analysis described in the major comments.

major comments (5)
  1. [§5, §6.1] The protocol description states that training is iterative over multiple FL rounds, but the experimental section does not report the number of FL rounds or local epochs, and the accuracy numbers in Section 6.1 appear to be from a single aggregation of three client models. Figure 3 then linearly extrapolates the per-round communication measurement to '10 rounds.' Because the paper's central claims about scalability and about FLHHE as a federated learning method depend on multi-round behavior, the authors should report the full training configuration (rounds, local epochs, learning rate, batch size) and provide multi-round experiments with a convergence curve.
  2. [§6, Computation Analysis] The server-side transciphering cost (6.75 s online plus 32.94 s offline per client, Table 1) is the main trade-off of the scheme, but the paper supports its scalability claim only with the assertion that the computation 'can be highly optimized, for example, using careful orchestrated parallelize computation and/or distributed systems.' No parallel or distributed server implementation, measurement, or resource model is provided. For tens or hundreds of clients, the serial online cost alone would be minutes to hours per round; the authors should either measure a parallel/distributed setting or replace the scalability claim with a statement of what remains to be demonstrated.
  3. [§4.1, Tables 2 and 4] The per-round communication comparison omits the one-time HE-encrypted symmetric key listed in Table 2 as 738.5 MB. Since Section 4.1 gives each client a unique symmetric key Ki, the initial setup would transfer or store n such encrypted keys (one per client). Over the 10 training rounds used in Figure 3, this one-time cost can outweigh the per-round bandwidth savings of HHE over HE, so the paper needs to state who transmits cKi (key dealer vs client), whether it is per-client or shared, and to include it in a total-communication analysis.
  4. [§6, Tables 1 and 3] All timings are reported only as averages over 10 runs, without standard deviations, confidence intervals, or per-run values, and the tables are internally inconsistent (Transcipher online is 6.87 s in Table 1 but 6.75 s in Table 3; HE aggregation is 0.004 s in Table 3 while the text says 0.003 s). Without variance information, the claimed 26x reduction in client encryption time (0.008 s vs 0.21 s) cannot be assessed for statistical robustness, and the inconsistent numbers should be reconciled.
  5. [§4.2, §5] The paper claims that FLHHE is 'secure' and lists threat-model requirements, but it does not provide a formal security definition, proof, or reduction to the underlying HHE scheme. In particular, it does not analyze what the server could infer from the aggregated model or whether the use of a trusted key dealer and a single round of aggregation changes the security guarantees. A precise security statement is needed for the title's 'secure' claim.
minor comments (5)
  1. [§3.1, §6] Section 3.1: 'let’s assume their are N users' should be 'there are'; Section 6: 'statically insignificant' should be 'statistically insignificant'; Section 2 has several formatting errors such as 'Madiet al.', 'Shiet al.,', and 'Stripelis et al.,,'.
  2. [§6.1] The section heading 'Research Question Answers' appears three times in a row; merge these into a single discussion to improve readability.
  3. [Figure 3] The caption and text should state explicitly that the curves are a linear extrapolation of the single-round, three-client measurement, not measured results at 5, 10, or 20 clients, and should note the assumed number of rounds and whether the one-time HE-encrypted key is included.
  4. [§6.1] Reporting identical accuracies for plaintext, HE, and HHE is a useful sanity check that the transciphering pipeline introduces no numerical error, but it should be described as a consistency check rather than as evidence about the quality of the federated learning process; the 65.92% full-test-set accuracy also reflects the label-exclusion partitioning and should be interpreted only as a toy demonstration.
  5. [§7, Remark 1] The claim that the method 'can be easily extended' to healthcare settings is not supported by any healthcare dataset or domain-specific experiment; consider presenting this as future work rather than an insight of the current study.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: FLHHE's communication, computation, and accuracy results are measured from a concrete implementation, not derived from its own inputs.

full rationale

The paper's quantitative claims—client upload of 1.048 MB vs. 16.8 MB, client-side encryption of 0.008 s vs. 0.21 s, server-side transciphering of 6.75 s, and equal accuracies across plaintext/HE/HHE—are all reported measurements from a concrete implementation (PyTorch, Lattigo, Rubato/RtF) described in Section 6. The accuracy equality is a functional-correctness check: FLHHE and HE are defined to compute the same SimpAvg over the same decrypted local weights, so equal accuracy is expected and confirms correct transciphering/decryption; it is not a prediction obtained by fitting. No parameter is fitted to a target result, and no 'prediction' coincides by construction with an input. The protocol adopts Rubato/RtF from external prior work [6,18], while the self-citations [1,14,20,23] appear only in background statements about HE cost and HHE design, not as load-bearing uniqueness or existence arguments. The concern about a single-round, 3-client simulation and Figure 3's linear extrapolation is an external-validity/correctness issue, not circularity. Therefore no circular step is identified.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central contribution is an empirical integration; it introduces no new cryptographic primitive, so the ledger is mostly configuration choices and inherited trust assumptions. No free parameters are fitted to make accuracy match; the cost figures are direct measurements. The main uncharged assumptions are the correctness and security of Rubato/RtF and the extrapolation from a toy setup to real deployments.

free parameters (3)
  • Training hyperparameters (learning rate, batch size, local epochs) = not reported
    Accuracy results depend on these choices, but the paper omits them, making the accuracy claim hard to reproduce exactly.
  • Data partition and client count = 3 clients; partitions exclude labels (1,3,7), (2,5,8), (4,6,9)
    The non-IID partition with mutually excluded labels determines the low 65.92 percent accuracy and is a configuration choice, not a property of FLHHE.
  • Neural network architecture = 784-32-10 two-layer MLP, no bias
    Cost and accuracy numbers are specific to this tiny model; scaling behavior to larger models is not measured.
assumptions (3)
  • domain assumption Rubato and RtF transciphering are correct and provide the claimed security level.
    The protocol inherits correctness and security from the cited framework [18]; no independent proof or cryptanalysis is provided. Invoked in Sections 3.4 and 4.2 as the basis for privacy.
  • domain assumption Semi-honest clients and server, and a trusted key dealer that does not collude.
    Section 4.2 defines this threat model. All privacy conclusions are conditional on these trust assumptions.
  • domain assumption CKKS and FV approximate arithmetic preserves model weights closely enough that accuracy is unaffected for this model.
    Section 6.1 reports identical accuracies but no weight-level comparison; this may fail for deeper models, more clients, or more aggregation rounds.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Privacy-Centric Approach: Scalable and Secure Federated Learning Enabled by Hybrid Homomorphic Encryption." pith.science (2026). https://pith.science/paper/F4OE2FWY

@misc{pith2026250714853,
  author       = {Pith},
  title        = {Pith review of: A Privacy-Centric Approach: Scalable and Secure Federated Learning Enabled by Hybrid Homomorphic Encryption},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F4OE2FWY}},
  note         = {Machine review of arXiv:2507.14853}
}
read the original abstract

Federated Learning (FL) enables collaborative model training without sharing raw data, making it a promising approach for privacy-sensitive domains. Despite its potential, FL faces significant challenges, particularly in terms of communication overhead and data privacy. Privacy-preserving Techniques (PPTs) such as Homomorphic Encryption (HE) have been used to mitigate these concerns. However, these techniques introduce substantial computational and communication costs, limiting their practical deployment. In this work, we explore how Hybrid Homomorphic Encryption (HHE), a cryptographic protocol that combines symmetric encryption with HE, can be effectively integrated with FL to address both communication and privacy challenges, paving the way for scalable and secure decentralized learning system.

Figures

Figures reproduced from arXiv: 2507.14853 by the authors.

Figure 1
Figure 1. The RtF transciphering framework. symmetric cipher E, the client computes the FV-ciphertext K of the symmetric secret key k and sends it to the server. Client-side Computation: – Offline Phase: The client generates a keystream z ∈ Z n q from the cipher E, taking the nonce nc ∈ {0, 1} λ and the secret key k ∈ Z n q as inputs. – Online Phase: First, the client scales up the plaintext message m (in floating point) by a… view at source ↗
Figure 2
Figure 2. FL Framework with HHE for Secure Model Aggregation [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Communication cost vs. number of clients save a lot of upload cost compared to HE (16 times less) by sending a symmetric encrypted model instead of an HE encrypted one. Nevertheless, since the FL server returns the same HE-encrypted aggregated model in both cases, the total communication cost of FLHHE for each client is reduced only 2 times, compared to HE. However, consider the fact that in real-world FL deployment… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 38 canonical work pages

  1. [1]

    In: International Conference on Security and Privacy in Communication Systems

    Bakas, A., Frimpong, E., Michalas, A.: Symmetrical disguise: Realizing homomor- phic encryption services from symmetric primitives. In: International Conference on Security and Privacy in Communication Systems. pp. 353–370. Springer (2022)

    work page 2022

  2. [2]

    In: International conference on security and privacy in communication systems

    Bakas, A., Michalas, A.: Modern family: A revocable hybrid encryption scheme based on attribute-based encryption, symmetric searchable encryption and sgx. In: International conference on security and privacy in communication systems. pp. 472–486. Springer (2019)

    work page 2019

  3. [3]

    Journal of Cryptology31(3), 885–916 (2018)

    Canteaut, A., Carpov, S., Fontaine, C., Lepoint, T., Naya-Plasencia, M., Paillier, P., Sirdey, R.: Stream ciphers: A practical solution for efficient homomorphic-ciphertext compression. Journal of Cryptology31(3), 885–916 (2018)

    work page 2018

  4. [4]

    In: International Conference on the Theory and Application of Cryptology and Information Security

    Cheon, J.H., Kim, A., Kim, M., Song, Y.: Homomorphic encryption for arithmetic of approximate numbers. In: International Conference on the Theory and Application of Cryptology and Information Security. pp. 409–437. Springer (2017)

    work page 2017

  5. [5]

    In: Advances in Cryptology– ASIACRYPT 2016

    Chillotti, I., Gama, N., Georgieva, M., Izabachene, M.: Faster fully homomorphic encryption: Bootstrapping in less than 0.1 seconds. In: Advances in Cryptology– ASIACRYPT 2016. pp. 3–33. Springer (2016)

    work page 2016

  6. [6]

    In: International Conference on the Theory and Application of Cryptology and Information Security

    Cho, J., Ha, J., Kim, S., Lee, B., Lee, J., Lee, J., Moon, D., Yoon, H.: Transciphering framework for approximate homomorphic encryption. In: International Conference on the Theory and Application of Cryptology and Information Security. pp. 640–669. Springer (2021) 2 https://github.com/khoaguin/flhhe A Privacy-Centric Approach 15

    work page 2021

  7. [7]

    Clet, P.E., Stan, O., Zuber, M.: Bfv, ckks, tfhe: Which one is the best for a secure neural network evaluation in the cloud? In: Applied Cryptography and Network Security Workshops: ACNS 2021 Satellite Workshops. pp. 279–300. Springer (2021)

    work page 2021

  8. [8]

    Cosseron, O., Hoffmann, C., Méaux, P., Standaert, F.X.: Towards case-optimized hybridhomomorphicencryption:Featuringtheelisabethstreamcipher.In:Advances in Cryptology – ASIACRYPT 2022. p. 32–67. Springer-Verlag (2023)

    work page 2023

Show all 40 references
  1. [9]

    Transaction on Cryptographic Hardware and Embedded Systems 2023 Issue 3 (2023)

    Dobraunig, C., Grassi, L., Helminger, L., Rechberger, C., Schofnegger, M., Walch, R.: Pasta: a case for hybrid homomorphic encryption. Transaction on Cryptographic Hardware and Embedded Systems 2023 Issue 3 (2023)

    2023

  2. [10]

    Cryp- tology ePrint Archive (2012)

    Fan, J., Vercauteren, F.: Somewhat practical fully homomorphic encryption. Cryp- tology ePrint Archive (2012)

    2012

  3. [11]

    Future Internet13(4), 94 (2021)

    Fang, H., Qian, Q.: Privacy preserving machine learning with homomorphic encryp- tion and federated learning. Future Internet13(4), 94 (2021)

    2021

  4. [12]

    Journal of Intelligent Learning Systems and Applications p

    Fatima, M., Pasha, M., et al.: Survey of machine learning algorithms for disease diagnostic. Journal of Intelligent Learning Systems and Applications p. 1 (2017)

    2017

  5. [13]

    In: Proceedings of the 22nd ACM SIGSAC conference on computer and communications security

    Fredrikson, M., Jha, S., Ristenpart, T.: Model inversion attacks that exploit confi- dence information and basic countermeasures. In: Proceedings of the 22nd ACM SIGSAC conference on computer and communications security. pp. 1322–1333 (2015)

    2015

  6. [14]

    39th ACM/SIGAPP Symposium On Applied Computing (SAC’24) (2024)

    Frimpong, E., Nguyen, K., Budzys, M., Khan, T., Michalas, A.: Guardml: Effi- cient privacy-preserving machine learning services through hybrid homomorphic encryption. 39th ACM/SIGAPP Symposium On Applied Computing (SAC’24) (2024)

    2024

  7. [15]

    Geiping, J., Bauermeister, H., Dröge, H., Moeller, M.: Inverting gradients-how easy is it to break privacy in federated learning? Advances in neural information processing systems 33, 16937–16947 (2020)

    2020

  8. [16]

    Stanford university (2009)

    Gentry, C.: A fully homomorphic encryption scheme. Stanford university (2009)

    2009

  9. [17]

    In: Annual Cryptology Conference

    Gentry, C., Halevi, S., Smart, N.P.: Homomorphic evaluation of the aes circuit. In: Annual Cryptology Conference. Springer (2012)

    2012

  10. [18]

    In: Advances in Cryptology–EUROCRYPT 2022

    Ha, J., Kim, S., Lee, B., Lee, J., Son, M.: Rubato: Noisy ciphers for approximate homomorphic encryption. In: Advances in Cryptology–EUROCRYPT 2022. pp. 581–610. Springer (2022)

    2022

  11. [19]

    In: Proceedings of the 29th ACM Symposium on Access Control Models and Technologies

    Khan, T., Budzys, M., Michalas, A.: Make split, not hijack: Preventing feature-space hijacking attacks in split learning. In: Proceedings of the 29th ACM Symposium on Access Control Models and Technologies. pp. 19–30 (2024)

    2024

  12. [20]

    arXiv preprint arXiv:2507.10494 (2025)

    Khan, T., Budzys, M., Michalas, A.: Split happens: Combating advanced threats with split learning and function secret sharing. arXiv preprint arXiv:2507.10494 (2025)

    2025 arXiv

  13. [21]

    In: Proceedings of the 24th Privacy Enhancing Technologies Symposium (PETS’24)

    Khan, T., Budzys, M., Nguyen, K., Michalas, A.: Sok: Wildest dreams: Reproducible research in privacy-preserving neural network training. In: Proceedings of the 24th Privacy Enhancing Technologies Symposium (PETS’24). Springer-Verlag, Berlin, Heidelberg (2024)

    2024

  14. [22]

    In: 2023 IEEE 22nd International Conference on Trust, Security and Privacy in Computing and Communications (TrustCom)

    Khan, T., Michalas, A.: Learning in the dark: Privacy-preserving machine learning using function approximation. In: 2023 IEEE 22nd International Conference on Trust, Security and Privacy in Computing and Communications (TrustCom). pp. 62–71. IEEE (2023)

    2023

  15. [23]

    In: 2023 20th Annual International Conference on Privacy, Security and Trust (PST)

    Khan, T., Nguyen, K., Michalas, A., Bakas, A.: Love or hate? share or split? privacy-preserving training using split learning and homomorphic encryption. In: 2023 20th Annual International Conference on Privacy, Security and Trust (PST). pp. 1–7. IEEE (2023) 16 K. Nguyen et al

    2023

  16. [24]

    arXiv preprint arXiv:1610.05492 (2016)

    Konečn` y, J., McMahan, H.B., Yu, F.X., Richtárik, P., Suresh, A.T., Bacon, D.: Fed- erated learning: Strategies for improving communication efficiency. arXiv preprint arXiv:1610.05492 (2016)

    2016 arXiv

  17. [25]

    IEEE Access10, 30039–30054 (2022)

    Lee, J.W., Kang, H., Lee, Y., Choi, W., Eom, J., Deryabin, M., Lee, E., Lee, J., Yoo, D., Kim, Y.S., et al.: Privacy-preserving machine learning with fully homomorphic encryption for deep neural network. IEEE Access10, 30039–30054 (2022)

    2022

  18. [26]

    IEEE Transactions on Knowledge and Data Engineering35(4), 3347–3366 (2021)

    Li, Q., Wen, Z., Wu, Z., Hu, S., Wang, N., Li, Y., Liu, X., He, B.: A survey on federated learning systems: Vision, hype and reality for data privacy and protection. IEEE Transactions on Knowledge and Data Engineering35(4), 3347–3366 (2021)

    2021

  19. [27]

    In: 2021 Reconciling Data Analytics, Automation, Privacy, and Security: A Big Data Challenge (RDAAPS)

    Madi, A., Stan, O., Mayoue, A., Grivet-Sébert, A., Gouy-Pailler, C., Sirdey, R.: A secure federated learning framework using homomorphic encryption and verifiable computing. In: 2021 Reconciling Data Analytics, Automation, Privacy, and Security: A Big Data Challenge (RDAAPS). ...

    2021

  20. [28]

    In: Proceedings of the 2019 ACM SIGSAC Conference on Cloud Computing Security Workshop

    Mandal, K., Gong, G.: Privfl: Practical privacy-preserving federated regressions on high-dimensional data over mobile networks. In: Proceedings of the 2019 ACM SIGSAC Conference on Cloud Computing Security Workshop. pp. 57–68 (2019)

    2019

  21. [29]

    McMahan, H.B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data (2023), https://arxiv.org/abs/1602.05629

    2023 arXiv

  22. [30]

    In: International Conference on Cryptology in India

    Méaux, P., Carlet, C., Journault, A., Standaert, F.X.: Improved filter permutators for efficient fhe: Better instances and implementations. In: International Conference on Cryptology in India. Springer (2019)

    2019

  23. [31]

    In: Security and Privacy in Communication Networks (Se- cureComm’23), Hong Kong SAR, Hong Kong, 19—21 October 2023

    Nguyen, K., Khan, T., Michalas, A.: Split without a leak: Reducing privacy leak- age in split learning. In: Security and Privacy in Communication Networks (Se- cureComm’23), Hong Kong SAR, Hong Kong, 19—21 October 2023. Springer Nature Switzerland, Cham (2023)

    2023

  24. [32]

    In: 2022 13th International Conference on Information and Communication Technology Convergence (ICTC)

    Park, J., Yu, N.Y., Lim, H.: Privacy-preserving federated learning using homo- morphic encryption with different encryption keys. In: 2022 13th International Conference on Information and Communication Technology Convergence (ICTC). pp. 1869–1871. IEEE (2022)

    2022

  25. [33]

    In: The adaptive web, pp

    Pazzani, M.J., Billsus, D.: Content-based recommendation systems. In: The adaptive web, pp. 325–341. Springer (2007)

    2007

  26. [34]

    Proceedings on Privacy En- hancing Technologies2022(1) (2020)

    Ryffel, T., Tholoniat, P., Pointcheval, D., Bach, F.: Ariann: Low-interaction privacy- preserving deep learning via function secret sharing. Proceedings on Privacy En- hancing Technologies2022(1) (2020)

    2020

  27. [35]

    Telecommunication Systems 82(3), 419–433 (2023)

    Shi,Z.,Yang,Z.,Hassan,A.,Li,F.,Ding,X.:Aprivacypreservingfederatedlearning scheme using homomorphic encryption and secret sharing. Telecommunication Systems 82(3), 419–433 (2023)

    2023

  28. [36]

    In: 17th Interna- tional Symposium on Medical Information Processing and Analysis

    Stripelis, D., Saleem, H., Ghai, T., Dhinagar, N., Gupta, U., Anastasiou, C., Ver Steeg, G., Ravi, S., Naveed, M., Thompson, P.M., et al.: Secure neuroimaging analysis using federated learning with homomorphic encryption. In: 17th Interna- tional Symposium on Medical Informati...

    2021

  29. [37]

    In: 2022 IEEE Symposium on Security and Privacy (SP)

    Tian, H., Zeng, C., Ren, Z., Chai, D., Zhang, J., Chen, K., Yang, Q.: Sphinx: En- abling privacy-preserving online learning over the cloud. In: 2022 IEEE Symposium on Security and Privacy (SP). pp. 2487–2501. IEEE (2022)

    2022

  30. [38]

    Wagh, S., Gupta, D., Chandran, N.: Securenn: 3-party secure computation for neural network training. Proc. Priv. Enhancing Technol.2019(3), 26–49 (2019)

    2019

  31. [39]

    ACM Transactions on Intelligent Systems and Technology (TIST) 10(2), 1–19 (2019) A Privacy-Centric Approach 17

    Yang, Q., Liu, Y., Chen, T., Tong, Y.: Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems and Technology (TIST) 10(2), 1–19 (2019) A Privacy-Centric Approach 17

    2019

  32. [40]

    Advances in neural informa- tion processing systems32 (2019)

    Zhu, L., Liu, Z., Han, S.: Deep leakage from gradients. Advances in neural informa- tion processing systems32 (2019)

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.