Pith. sign in

REVIEW 3 major objections 3 minor 53 references

Information-Theoretically Secure Aggregation for Lightweight Federated Learning: Resilient to Dropouts and Adversaries

T0 review · 3 major / 3 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read A single masked opening suffices to compute the federated majority vote with information-theoretic privacy.

desk verdict The core reduction breaks at x=0 and Method B cannot evaluate negative powers, so the headline efficiency claims don't follow — but the single-base structural idea is worth salvaging. read the letter →

arxiv 2607.20890 v1 pith:NOFZVGQW submitted 2026-07-23 cs.LG

classification cs.LG
keywords information-theoreticsecurityfederatedlearningsignSGDmajorityvotepolynomialsecureaggregationsingle-roundmultiplicationdropoutresilienceMDScodes
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

The paper asks whether sign-based federated learning—where each device sends one bit, the sign of its gradient, and the server takes a majority vote—can be aggregated without revealing any individual sign. It answers yes under an honest-majority assumption, by viewing the majority vote as a polynomial in a single aggregated value and evaluating that polynomial with secure multiparty computation in a single round. The central construction uses one random mask: the devices open only the masked sum, from which the majority vote polynomial is reconstructed on secret-shared coefficients. If correct, this gives information-theoretic privacy (security against unbounded adversaries) rather than computational security, while cutting online communication by up to 99.5% and latency by 85.7% relative to Beaver-triple approaches. The paper also claims that the same MDS-code machinery that opens the masked value lets the server recover dropped users' contributions and correct corrupted shares, so accuracy is preserved even with 35.7% dropouts or 23.9% adversarial shares.

What carries the argument

The single-base structure of the majority-vote polynomial, combined with Fermat's Little Theorem. The MV polynomial is univariate in x = sum_i sign(g_i). Method B (Single-Mask Power Sharing) exploits this: for one random mask a, precompute secret-shared powers [a^j] and coefficient-weighted bundles [P_r] = sum_j c_j * binom(j,r) * [a^(j-r)]; in the online round, open only x_hat = x - a, then everyone computes [F(x)] = sum_r x_hat^r [P_r]. The binomial expansion turns one opening into all powers of x. Inverse-form exponent reduction replaces exponents j > (p-1)/2 with -(p-1-j), halving the polynomial degree. MDS-code-based decoding of the single opening is what provides dropout recovery and a

What would settle it

Pick n even, say n = 4 with p = 5, where the MV polynomial with tie-break sign(0) = 0 reduces to something containing x^(-1) after inverse-form reduction. Simulate the protocol on a round where two users send +1 and two send -1, so x = 0. The online phase opens x_hat = x - a, and the polynomial evaluation step must handle a negative exponent; when x = 0, x_hat = -a, so the term x^(-1) would require computing (-a)^(-1), which is undefined if a = 0 and invalid at x = 0 even if a ≠ 0. The concrete check is whether the protocol's reconstruction yields the tie-breaking value F(0) or fails to decode

Watch

Extended reading notes

Core claim

The paper's central claim is that the majority-vote polynomial F(x) = sum_m sign(m)(1-(x-m)^(p-1)) over F_p — the polynomial whose value at x is the majority sign of n one-bit gradients — can be evaluated securely in a single communication round. The key observation is that every monomial is a power of one base value x = sum_i sign(g_i), rather than a product of distinct inputs. Therefore, instead of generically multiplying many inputs, the protocol masks x once with a random value a and opens only x_hat = x - a; using the binomial expansion (x_hat + a)^j and precomputed secret shares of powers of a, each party computes a share of F(x) locally. The opened value is statistically independent o

Load-bearing premise

The whole efficiency gain relies on replacing high-degree terms x^j with x^(-(p-1-j)) using x^(p-1) ≡ 1, which holds only for nonzero x; at x = 0 — exactly a tied majority vote — the inverse is undefined, and the paper does not show how Method B evaluates negative powers of the masked value.

Editorial extensions

If this is right

  • Sign-based federated learning can get information-theoretic privacy with a single online field element per user, making secure aggregation practical for lightweight devices.
  • The protocol reduces online communication by up to 99.5% and latency by up to 85.7% versus Beaver-triple-based secure aggregation, while keeping offline storage linear in the polynomial degree rather than exponential.
  • Because the opening is MDS-coded, the server can recover the exact majority vote even when users drop out or send corrupted shares, up to the decoding bound 2e + s ≤ n - 2t - 1.
  • The server learns only the final majority sign; individual gradient signs are statistically hidden from the server and from any coalition of fewer than n/2 users.
  • The convergence guarantees of signSGD with majority vote are preserved whenever the error-erasure condition holds, because the decoded aggregate equals the benign majority vote.

Reading between the lines

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

  • The same single-mask polynomial evaluation extends beyond majority vote to any univariate function of the aggregated sum x — for example, top-k indicators, ternary thresholds, or clipped means — as long as it can be written as a polynomial over F_p of degree below p. That would give a general recipe for private aggregation of one-bit updates.
  • If the inverse-form reduction is restricted to inputs where x ≠ 0, the tie case must be handled separately; a testable fix is to split F(x) into a nonzero-domain part plus an explicit tie-breaking constant, so that no negative exponent is ever evaluated at zero.
  • Choosing a deliberately small privacy threshold t (e.g., n/4 instead of n/2) enlarges the MDS decoding region, trading some collusion resistance for much higher dropout and adversary tolerance — a tunable knob the paper quantifies but does not fully explore in its experiments.
  • The method could combine with hierarchical or clustered federated learning: if each subgroup's majority vote is a univariate polynomial, the same one-opening trick applies per subgroup, keeping per-subgroup communication constant and making the overhead scale gracefully with system size.
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

3 major / 3 minor

Summary. The paper proposes an information-theoretically secure aggregation framework for signSGD-based federated learning. It represents the majority-vote function as a polynomial over a prime field, applies an 'inverse-form exponent reduction' to lower the polynomial degree to about half, and evaluates the reduced polynomial by one of two single-round secure-multiplication methods: Method A (DN + N-BTE) or Method B (single-mask power sharing). The authors claim end-to-end privacy under an honest-majority assumption, resilience to dropouts and adversarial users via MDS decoding, and large reductions in online communication and latency. The core technical claim is that the inverse-form reduction preserves the majority-vote polynomial; the paper's central protocol is built on this reduction.

Significance. The paper identifies a genuinely useful structural property of MV polynomials — all monomials are powers of a single aggregated value — and the single-mask binomial-expansion approach (Method B) would be elegant and efficient for polynomials with only nonnegative exponents. The honest-majority, information-theoretic framing and the MDS-based robustness discussion in Section VI-A are coherent and could be valuable in a corrected design. However, the inverse-form exponent reduction is load-bearing for every efficiency claim in the abstract, Section V, and Table IV, and it is invalid at x=0 and cannot be evaluated by either proposed method. Since the central protocol as written does not compute the majority vote, the convergence and security theorems do not apply to the executed computation. The contribution, as stated, is not sound.

major comments (3)
  1. [Section V-A, Theorem 1] The identity x^j ≡ x^{-(p-1-j)} is asserted for all x∈F_p, with the proof saying it is 'trivial for x=0'. This is false: Fermat's Little Theorem gives x^{p-1}=1 only on F_p^×, and 0^{-e} is undefined in F_p. The point x=0 is not a corner case for this application because Section III-C defines tie-breaking for x=0 when n is even. Concretely, for n=4, p=5, Table II gives F(x)=x^4+3x^3+x+4, with F(0)=4 (sign(0)=-1); the inverse-form reduction in Section V-A rewrites this as 3x^{-1}+x, which is undefined at 0 (and equals 0 if one substitutes x^{-1}=x^{p-2}), not 4. The reduction therefore does not preserve F(x) on the domain required by the MV problem.
  2. [Section V-B2, Eq. (8), Algorithm 6] Method B evaluates only polynomials with nonnegative exponents. Eq. (8) expands (x̂+a)^k for k∈Z_{\ge0}; Algorithm 6 precomputes [a^j] and [P_r] for j,r=0..N', opens x̂=x-a, computes public powers x̂^0,...,x̂^{N'}, and forms [F(x)]_t = Σ_{r=0}^{N'} x̂^r[P_r]_t. There is no term anywhere that computes (x̂+a)^{-e} or x^{-e}. Thus for the n=4 reduced polynomial 3x^{-1}+x, the inverse term is silently omitted, because its exponent is not in {0,...,N'} with N'=2. The same holds for Method A in Section V-B1, which only multiplies formal copies x^k for k≥0. Negative exponents cannot be produced by the binomial expansion used.
  3. [Abstract, Section V, Table IV] The headline reductions (up to 99.5% online communication, 85.7% latency) and the linear offline complexity O(N') with N'=(p-1)/2 all depend on the inverse-form exponent reduction. Since that reduction is invalid at x=0 and cannot be evaluated by either secure-multiplication method, these claims are unsupported. The protocol as written does not compute the MV polynomial; consequently Theorem 3 and Theorem 4 are proved for a functionality that Algorithm 6 does not actually implement. The paper needs either a negative-exponent secure-evaluation mechanism that also handles x=0 correctly, or a return to the unreduced degree-(p-2) polynomial with a matching evaluation method.
minor comments (3)
  1. [Algorithm 5, step 14] The line '[z]t ← QN l=1 u_l {public value}' mixes a public value with t-sharings; the constant should be explicitly embedded, e.g., as [1]_t or by adding it to one party's share.
  2. [Section VI-B, Theorem 3] The notation n_t = W^2 is unexplained; W is not defined in the theorem statement, and the relationship between n_t and the active-user count n is unclear. The proof sketch also simply substitutes α=0 into [14] without a self-contained derivation.
  3. [Table I] The row for the proposed method lists 'Very High' scalability while Method A has exponential offline cost; the table would benefit from distinguishing Method A and Method B, since only Method B achieves the reported linear offline complexity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: core derivations are self-contained; self-citations are not load-bearing.

full rationale

The paper's central derivation chain is not circular. The MV polynomial is constructed from Fermat's Little Theorem in Section III-C, with Eq. (2) citing the authors' prior Hi-SAFE work but also providing an independent derivation via indicator functions. Method B's correctness is proven internally using the binomial theorem (Lemma 1 and Theorem 2), and its security is established through a standard simulation-based argument with a freshly sampled uniform mask; no parameter is fitted and no 'prediction' is obtained by renaming an input. The self-citations to [25] and [19] provide background, baselines, and a convenient polynomial formula, but they are not the sole justification for the central claims. The convergence analysis imports an external bound from [14] and substitutes α=0, which is a legitimate application of a known result rather than a self-referential derivation. The paper does contain a serious mathematical/operational flaw: the inverse-form exponent reduction of Theorem 1 is invalid at x=0 (where the tie-breaking case lives), and Method B/Algorithm 6 only computes nonnegative powers, so the reduced polynomial's inverse terms are not actually evaluated. This is a correctness risk, not circularity, and therefore does not affect the circularity score.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

The protocol rests on standard MPC assumptions (honest majority, semi-honest model, Shamir secret sharing, Reed-Solomon decoding) plus an unstated assumption that all users' sign gradients have already been secret-shared among the participants. No free parameters are fitted. The principal ad-hoc axiom is the claimed equivalence of the MV polynomial with its inverse-form reduced form on the entire field, including x=0.

assumptions (6)
  • standard math Fermat's Little Theorem: x^{p-1} ≡ 1 (mod p) for x ≠ 0
    Used in Eq. (1) and Theorem 1 to construct indicators and reduce polynomial degrees.
  • standard math Shamir t-sharing and DN multiplication are secure under t < n/2
    Basis for the proposed secure multiplication; standard MPC result.
  • standard math Reed-Solomon/MDS decoding of 2t-sharings corrects e errors and s erasures when 2e+s ≤ n-2t-1
    Used for dropout/adversary robustness in Section VI-A.
  • ad hoc to paper The inverse-form reduced polynomial F(x)=Σ c_j x^{-(p-1-j)} is equivalent to the original MV polynomial for all x∈F_p, including x=0
    Assumed in Section V-A; false at x=0 because the negative exponent is undefined and x^{p-1}≠1 at 0.
  • domain assumption Each user's sign gradient is already secret-shared among all n users before Algorithm 6 runs
    Method B Step 1 requires every party to hold shares of all inputs; the paper never specifies or counts this sharing step.
  • domain assumption Semi-honest model and honest majority t < n/2 align with signSGD-MV correctness
    Stated in Section IV-C; required for information-theoretic security.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Information-Theoretically Secure Aggregation for Lightweight Federated Learning: Resilient to Dropouts and Adversaries." pith.science (2026). https://pith.science/paper/NOFZVGQW

@misc{pith2026260720890,
  author       = {Pith},
  title        = {Pith review of: Information-Theoretically Secure Aggregation for Lightweight Federated Learning: Resilient to Dropouts and Adversaries},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NOFZVGQW}},
  note         = {Machine review of arXiv:2607.20890}
}
read the original abstract

On-device federated learning (FL) enables privacy-preserving and personalized model training on resource-constrained devices such as smartphones and IoT nodes. To reduce communication cost, sign-based methods (e.g., signSGD) transmit one-bit gradients. However, exposing gradient signs makes them vulnerable to inference attacks, while existing secure aggregation schemes are often incompatible with such methods or incur significant computational and communication overhead. We propose a lightweight and information-theoretically secure aggregation framework tailored for sign-based FL. The framework securely computes the majority vote (MV) polynomial through single-round secure multiplication, ensuring end-to-end information-theoretic security under the honest-majority assumption while revealing only the final aggregated sign to the server. To enhance efficiency and scalability, we introduce two key techniques. First, inverse-form exponent reduction halves the effective MV polynomial degree, reducing both communication and computation costs. Second, we propose single-round secure multiplication, achieving linear offline complexity and storage with only a single online communication. Together, these techniques reduce online communication by up to 99.5% and latency by up to 85.7% compared to conventional approaches. Also, by leveraging inherent MDS-code-based decoding, the framework achieves robustness against both dropouts and adversarial behaviors, yielding accuracy gains of up to 20.65% and 10.74%, respectively. Overall, the proposed framework establishes a practical foundation for large-scale, low-latency, and information-theoretically secure aggregation in sign-based FL.

Figures

Figures reproduced from arXiv: 2607.20890 by the authors.

Figure 1
Figure 1. Overview of the proposed FL framework. Sign gradi [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Performance comparison under different dropout rates [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 2
Figure 2. Comparison of online performance: (a) communication [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Performance comparison under different adversarial [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Performance comparison under different adversarial [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 9 linked inside Pith

  1. [1]

    Communication-efficient learning of deep networks from decentralized data,

    B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” inProceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS). PMLR, 2017, pp. 1273–1282

  2. [2]

    Federated learning: Challenges, methods, and future directions,

    T. Li, A. K. Sahu, A. Talwalkar, and V . Smith, “Federated learning: Challenges, methods, and future directions,”IEEE Signal Processing Magazine, vol. 37, no. 3, pp. 50–60, 2020

  3. [3]

    Toward on-device federated learning: A direct acyclic graph-based blockchain approach,

    M. Cao, L. Zhang, and B. Cao, “Toward on-device federated learning: A direct acyclic graph-based blockchain approach,”IEEE Transactions on Neural Networks and Learning Systems, pp. 2028–2042, Aug. 2021

  4. [4]

    Fedaux: Leveraging unlabeled auxiliary data in federated learning,

    F. Sattler, T. Korjakow, R. Rischke, and W. Samek, “Fedaux: Leveraging unlabeled auxiliary data in federated learning,”IEEE Transactions on Neural Networks and Learning Systems, pp. 5531–5543, Sep. 2023

  5. [5]

    Active client selection for clustered federated learning,

    H. Huang, W. Shi, Y . Feng, C. Niu, G. Cheng, J. Huang, and Z. Liu, “Active client selection for clustered federated learning,”IEEE Trans- actions on Neural Networks and Learning Systems, pp. 1–15, Jul. 2023

  6. [6]

    Federated learning with taskonomy for non-iid data,

    H. Jamali-Rad, M. Abdizadeh, and A. Singh, “Federated learning with taskonomy for non-iid data,”IEEE Transactions on Neural Networks and Learning Systems, vol. 34, no. 11, pp. 8719–8730, Mar. 2022

  7. [7]

    Practical and robust federated learning with highly scalable regression training,

    S. Han, H. Ding, S. Zhao, S. Ren, Z. Wang, J. Lin, and S. Zhou, “Practical and robust federated learning with highly scalable regression training,”IEEE Transactions on Neural Networks and Learning Systems, pp. 1–15, Jun. 2023

  8. [8]

    Personalized federated graph learning on non-iid electronic health records,

    T. Tang, Z. Han, Z. Cai, S. Yu, X. Zhou, T. Oseni, and S. K. Das, “Personalized federated graph learning on non-iid electronic health records,”IEEE Transactions on Neural Networks and Learning Systems, pp. 1–14, Mar. 2024

Show all 53 references
  1. [9]

    Clustered federated learning in het- erogeneous environment,

    Y . Yan, X. Tong, and S. Wang, “Clustered federated learning in het- erogeneous environment,”IEEE Transactions on Neural Networks and Learning Systems, pp. 1–14, Apr. 2023

  2. [10]

    Communication-efficient randomized algorithm for multi-kernel online federated learning,

    S. Hong and J. Chae, “Communication-efficient randomized algorithm for multi-kernel online federated learning,”IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 12, pp. 9872–9886, 2021

  3. [11]

    Tighter regret analysis and optimization of online federated learning,

    D. Kwon, J. Park, and S. Hong, “Tighter regret analysis and optimization of online federated learning,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

  4. [12]

    Federated learning in mobile edge networks: A comprehensive survey,

    W. Y . B. Lim, N. C. Luong, D. T. Hoang, Y . Jiao, Y .-C. Liang, Q. Yang, D. Niyato, and C. Miao, “Federated learning in mobile edge networks: A comprehensive survey,”IEEE communications surveys & tutorials, vol. 22, no. 3, pp. 2031–2063, 2020

  5. [13]

    Advances and open problems in federated learning,

    P. Kairouz, H. B. McMahan, B. Avent, A. Bellet, M. Bennis, A. N. Bhagoji, K. Bonawitz, Z. Charles, G. Cormode, R. Cummingset al., “Advances and open problems in federated learning,”Foundations and Trends in Machine Learning, vol. 14, no. 1–2, pp. 1–210, 2021

  6. [14]

    signSGD: Compressed optimisation for non-convex problems,

    J. Bernstein, Y .-X. Wang, K. Azizzadenesheli, and A. Anandkumar, “signSGD: Compressed optimisation for non-convex problems,” inPro- ceedings of the 35th International Conference on Machine Learning (ICML), 2018, pp. 560–569

  7. [15]

    1-bit stochastic gradient descent and its application to data-parallel distributed training of speech dnns,

    F. Seide, H. Fu, J. Droppo, G. Li, and D. Yu, “1-bit stochastic gradient descent and its application to data-parallel distributed training of speech dnns,” inFifteenth annual conference of the international speech communication association, 2014

  8. [16]

    signsgd with majority vote is communication efficient and fault tolerant,

    J. Bernstein, J. Zhao, K. Azizzadenesheli, and A. Anandkumar, “signsgd with majority vote is communication efficient and fault tolerant,”arXiv preprint arXiv:1810.05291, 2018

  9. [17]

    Sparse-signsgd with majority vote for communication-efficient distributed learning,

    C. Park and N. Lee, “Sparse-signsgd with majority vote for communication-efficient distributed learning,”arXiv preprint arXiv:2302.07475, 2023

  10. [18]

    Sign-based gradient descent with heterogeneous data: Convergence and byzantine resilience,

    R. Jin, Y . Liu, Y . Huang, X. He, T. Wu, and H. Dai, “Sign-based gradient descent with heterogeneous data: Convergence and byzantine resilience,” IEEE Transactions on Neural Networks and Learning Systems, pp. 1–13, Jan. 2024

  11. [19]

    FedLSC: Improving communi- cation efficiency and robustness in federated learning with stragglers and adversaries,

    H.-G. Joo, S. Hong, and D.-J. Shin, “FedLSC: Improving communi- cation efficiency and robustness in federated learning with stragglers and adversaries,”IEEE Transactions on Neural Networks and Learning Systems, vol. 36, no. 11, pp. 19 805–19 819, 2025

  12. [20]

    Exploiting unintended feature leakage in collaborative learning,

    L. Melis, C. Song, E. De Cristofaro, and V . Shmatikov, “Exploiting unintended feature leakage in collaborative learning,”Proceedings of IEEE Symposium on Security and Privacy (SP), pp. 691–706, 2019

  13. [21]

    Deep leakage from gradients,

    L. Zhu, Z. Liu, and S. Han, “Deep leakage from gradients,” inAdvances in Neural Information Processing Systems (NeurIPS), 2019, pp. 14 774– 14 784

  14. [22]

    Inverting gradients-how easy is it to break privacy in federated learning?

    J. Geiping, H. Bauermeister, H. Dr ¨oge, and M. Moeller, “Inverting gradients-how easy is it to break privacy in federated learning?”Ad- vances in neural information processing systems, vol. 33, pp. 16 937– 16 947, 2020

  15. [23]

    Practical secure aggregation for privacy-preserving machine learning,

    K. Bonawitz, V . Ivanov, B. Kreuter, A. Marcedone, H. B. McMahan, S. Patel, D. Ramage, A. Segal, and K. Seth, “Practical secure aggregation for privacy-preserving machine learning,” inProceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security (CCS), ...

  16. [24]

    Secure single-server aggregation with fault tolerance,

    J. Bell, A. Gasc ´on, M. Riazi, K. Singh, S. Su, and X. Wang, “Secure single-server aggregation with fault tolerance,” inProceedings of the 2020 ACM SIGSAC Conference on Computer and Communications Security (CCS), 2020, pp. 1169–1186

  17. [25]

    Hi-SAFE: Hierarchical secure aggregation for lightweight federated learning,

    H.-G. Joo, S. Hong, S. Lee, and D.-J. Shin, “Hi-SAFE: Hierarchical secure aggregation for lightweight federated learning,”IEEE Internet of Things Journal, 2026, https://doi.org/10.1109/JIOT.2026.3714526

  18. [26]

    Scalable and unconditionally secure multiparty computation,

    I. Damg ˚ard and J. B. Nielsen, “Scalable and unconditionally secure multiparty computation,” inAdvances in Cryptology–CRYPTO 2007. Springer, 2007, pp. 572–590

  19. [27]

    Turbo-aggregate: Breaking the quadratic aggregation barrier in secure federated learning,

    J. So, E. Shiet al., “Turbo-aggregate: Breaking the quadratic aggregation barrier in secure federated learning,” inProceedings of the 30th USENIX Security Symposium, 2021

  20. [28]

    A hybrid approach to privacy-preserving federated learning,

    S. Truex, N. Baracaldo, A. Anwar, T. Steinke, H. Ludwig, R. Zhang, and Y . Zhou, “A hybrid approach to privacy-preserving federated learning,” inProceedings of the 12th ACM workshop on artificial intelligence and security, 2019, pp. 1–11

  21. [29]

    Differentially private secure multi- party computation for federated learning in financial applications,

    D. Byrd and A. Polychroniadou, “Differentially private secure multi- party computation for federated learning in financial applications,” inProceedings of the First ACM International Conference on AI in Finance, 2020, pp. 1–9

  22. [30]

    DP-SIGNSGD: When efficiency meets privacy and robustness,

    L. Lyu, “DP-SIGNSGD: When efficiency meets privacy and robustness,” arXiv preprint arXiv:2105.04808, 2021

  23. [31]

    BatchCrypt: Efficient homomorphic encryption for cross-silo federated learning,

    C. Zhang, S. Li, J. Xia, W. Wang, F. Yan, and Y . Liu, “BatchCrypt: Efficient homomorphic encryption for cross-silo federated learning,” in 2020 USENIX annual technical conference (USENIX ATC 20), 2020, pp. 493–506

  24. [32]

    Privacy preserving machine learning with ho- momorphic encryption and federated learning,

    H. Fang and Q. Qian, “Privacy preserving machine learning with ho- momorphic encryption and federated learning,”Future Internet, vol. 13, no. 4, p. 94, 2021

  25. [33]

    FLASHE: Additively symmetric ho- momorphic encryption for cross-silo federated learning,

    Z. Jiang, W. Wang, and Y . Liu, “FLASHE: Additively symmetric ho- momorphic encryption for cross-silo federated learning,”arXiv preprint arXiv:2109.00675, 2021

  26. [34]

    Privacy-preserving federated learning based on multi-key homomorphic encryption,

    J. Ma, S.-A. Naas, S. Sigg, and X. Lyu, “Privacy-preserving federated learning based on multi-key homomorphic encryption,”International Journal of Intelligent Systems, vol. 37, no. 9, pp. 5880–5901, 2022

  27. [35]

    Homomorphic encryption for arithmetic of approximate numbers,

    J. H. Cheon, A. Kim, M. Kim, and Y . Song, “Homomorphic encryption for arithmetic of approximate numbers,” inAdvances in Cryptology– ASIACRYPT 2017: 23rd International Conference on the Theory and Applications of Cryptology and Information Security. Springer, 2017, pp. 409–437

  28. [36]

    Efficient fully homomorphic encryption from (standard) lwe,

    Z. Brakerski and V . Vaikuntanathan, “Efficient fully homomorphic encryption from (standard) lwe,”SIAM Journal on Computing, vol. 43, no. 2, pp. 831–871, 2014

  29. [37]

    Fully homomorphic encryption using ideal lattices,

    C. Gentry, “Fully homomorphic encryption using ideal lattices,”Pro- ceedings of the 41st annual ACM symposium on Theory of computing, pp. 169–178, 2009

  30. [38]

    Efficient multiparty protocols using circuit randomization,

    D. Beaver, “Efficient multiparty protocols using circuit randomization,” inAdvances in Cryptology—CRYPTO’91: Proceedings 11. Springer, 1992, pp. 420–432

  31. [39]

    Communication-efficient (client-aided) secure two-party protocols and its application,

    S. Ohata and K. Nuida, “Communication-efficient (client-aided) secure two-party protocols and its application,”arXiv preprint arXiv:1907.03415, 2020

  32. [40]

    ATLAS: efficient and scalable mpc in the honest majority setting,

    V . Goyal, H. Li, R. Ostrovsky, A. Polychroniadou, and Y . Song, “ATLAS: efficient and scalable mpc in the honest majority setting,” in Annual International Cryptology Conference. Springer, 2021, pp. 244– 274

  33. [41]

    Applications of fermat’s little theorem in cryptography,

    J. Smith, “Applications of fermat’s little theorem in cryptography,” Journal of Theoretical Cryptography, 2020

  34. [42]

    Communication-efficient learning of deep networks from decentralized 15 data,

    H. B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized 15 data,” inProceedings of the 20th International Conference on Artificial Intelligence and Statistics (AISTATS), 2017, pp. 1273–1282

  35. [43]

    Goldreich,Foundations of Cryptography: Volume 2–Basic Applica- tions

    O. Goldreich,Foundations of Cryptography: Volume 2–Basic Applica- tions. Cambridge University Press, 2004

  36. [44]

    Secure multiparty computation for privacy- preserving data mining,

    Y . Lindell and B. Pinkas, “Secure multiparty computation for privacy- preserving data mining,”Journal of Privacy and Confidentiality, vol. 1, no. 1, pp. 59–98, 2009

  37. [45]

    AHSecAgg and TSKG: Lightweight secure aggregation for federated learning without compromise,

    S. Zhang, Y . Liao, and P. Zhou, “AHSecAgg and TSKG: Lightweight secure aggregation for federated learning without compromise,”arXiv preprint arXiv:2312.04937, 2023

  38. [46]

    Secure aggregation of semi-honest clients and servers in federated learning with secret-shared homomorphism,

    D. Zhao, “Secure aggregation of semi-honest clients and servers in federated learning with secret-shared homomorphism,”ArXiv preprint arXiv:2303.10123, 2023

  39. [47]

    PQSF: Post-quantum secure privacy-preserving federated learning,

    H. Jiang, Y . Wang, Q. Luo, and X. Zhang, “PQSF: Post-quantum secure privacy-preserving federated learning,”Scientific Reports, vol. 14, no. 1, pp. 1–15, 2024

  40. [48]

    Secure and flexible privacy- preserving federated learning based on multi-key fully homomorphic encryption,

    C. Liu, Y . Wang, X. Chen, and H. Tan, “Secure and flexible privacy- preserving federated learning based on multi-key fully homomorphic encryption,”Electronics, vol. 13, no. 3, p. 567, 2024

  41. [49]

    von zur Gathen and J

    J. von zur Gathen and J. Gerhard,Modern Computer Algebra, 3rd ed. Cambridge University Press, 2013

  42. [50]

    Faster polynomial multiplication via multipoint kronecker substitution,

    D. Harvey, “Faster polynomial multiplication via multipoint kronecker substitution,” 2009. [Online]. Available: https://arxiv.org/abs/0912.0352

  43. [51]

    Gradient-based learning applied to document recognition,

    Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,”Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324, 1998

  44. [52]

    Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,

    H. Xiao, K. Rasul, and R. V ollgraf, “Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,”arXiv preprint arXiv:1708.07747, 2017

  45. [53]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, G. Hintonet al., “Learning multiple layers of features from tiny images,” 2009

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.