acceptodds
Under review as a conference paper at ICLR 2027

Efficient Deterministic Truncation and RMFE-Native Fixed-Point Operators for Private Neural Network Training

Abstract

Fixed-point arithmetic is widely used in privacy-preserving neural network training, but securely restoring the target scale of intermediate values remains costly. In the semi-honest, honest-majority setting with general -party Shamir sharing, existing mask-and-reveal deterministic truncation approaches either introduce a statistical security parameter to prevent modular wraparound or explicitly correct wraparound at the cost of sequential secure comparisons. In this work, we develop a more efficient deterministic truncation protocol for Shamir secret sharing over Galois rings. We revisit modular wraparound in the mask-and-reveal paradigm and incorporate its correction directly into the truncation relation. For Galois rings of characteristic , the formulation further simplifies, yielding a -free deterministic truncation protocol that requires only one online secure comparison. We further integrate this truncation protocol into the RMFE-based training framework and develop new protocols for efficient fixed-point multiplication and gradient aggregation, two key components of RMFE-based private neural network training for general parties. Experiments on fully connected and convolutional neural networks demonstrate up to a speedup in the LAN setting, together with substantial communication savings over the existing RMFE-based training framework.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
privacy-preserving machine learning, secure multi-party computation, truncation, fixed-point arithmetic, secret sharing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3QWXOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a κ-free deterministic truncation protocol for n-party semi-honest honest-majority MPC with Shamir secret-sharing over Galois rings. By directly integrating modular wraparound correction into the truncation formula, the protocol removes the statistical security parameter κ and only requires one online secure comparison, avoiding the downsides of two existing baselines: oversized rings from RMFE-style protocols and two sequential comparisons from Escudero-style κ-free schemes. Benchmarks on primitive micro-benchmarks and private neural-network training over MNIST demonstrate improved runtime and communication overhead compared with prior RMFE framework.

Critical strengths
  1. It achieves a favorable theoretical trade‑off for deterministic truncation under Galois-ring Shamir sharing: eliminating statistical security parameter κ while keeping only one online secure comparison, resolving two main bottlenecks of prior deterministic truncation constructions.
  2. The work is not limited to a standalone truncation primitive. It fully adapts the new truncation for RMFE SIMD-packed computation, and designs complete supporting operators including packed fixed‑point multiplication and gradient aggregation that can plug‑and‑play into the existing RMFE training pipeline.
Critical weaknesses
  1. The contribution of this paper is limited. Shamir sharing and truncation have been vastly studied in PPML.
  2. Although the paper provides rigorous theoretical derivation and end-to-end implementation, several practical boundary assumptions and experimental validations remain insufficiently explored.
  3. All end-to-end private neural network training experiments are only performed on MNIST. It remains unclear whether the observed performance gains and numerical stability hold for more complex datasets.
  4. Comprehensive hyperparameter sweeps are absent. The paper lacks full comparisons across varied packing density and multiple ring width combinations, so the parameter sensitivity of the proposed RMFE-native operators is not well characterized.
  5. It would be helpful to discuss the feasibility and overhead penalty when adapting ΠDTrun\Pi_{DTrun} and its supporting operators to malicious security MPC settings.
Questions for authors
  1. All end-to-end private training experiments are restricted to the MNIST dataset. Could the authors report key training results on at least one extra dataset such as CIFAR-10, to rule out the risk that reported performance gains are artefacts specific to MNIST?
  2. The protocol relies on strict ring-width constraint with only one-bit margin between input bit-length and ring bit-width. In practical DNN training, intermediate activations and gradients may occasionally exceed pre-estimated bit-length bounds. Could the authors quantify correctness or accuracy degradation when this one-bit margin constraint is violated?
  3. Could the authors supply additional experimental results over a broader range of hyperparameters, including diverse packing density and multiple ring width combinations, to better demonstrate parameter sensitivity of the proposed RMFE-native operators?
2 comments

We sincerely thank the reviewer for their valuable time and thoughtful feedback. Below, we address the key concerns raised in the review.

Regarding technical novelty, our contributions lie in two aspects. First, we derive a unified exact truncation relation by directly characterizing modular wraparound over a broader class of moduli NN. When specialized to Z2k\mathbb{Z}_{2^k}, this formulation enables κ\kappa-free deterministic truncation with only one online secure comparison. Second, we develop two RMFE-native operators specifically tailored to private neural network training: ΠMulRescale\Pi_{\mathrm{MulRescale}} for fixed-point multiplication and ΠAggRescale\Pi_{\mathrm{AggRescale}} for gradient aggregation. In particular, ΠMulRescale\Pi_{\mathrm{MulRescale}} introduces a novel full-ring random masking construction that enables direct truncation of degree-2t2t intermediates without separate degree reduction. Regarding malicious security, established MPC techniques could be employed to extend our constructions. This work focuses on establishing an efficient foundational framework in the semi-honest, honest-majority setting.

Q1 & Q3. We would like to emphasize that our primary contributions lie in fundamental MPC primitives, namely efficient deterministic truncation and RMFE-native fixed-point operators, rather than neural network architectures or training methodologies. The MNIST experiments serve to demonstrate the practical integration and efficiency of these primitives in private neural network training. Their correctness and security are established theoretically under the stated assumptions and do not depend on specific datasets or network architectures. Furthermore, we evaluate protocol-level performance and scalability across 3 to 15 parties. While larger datasets would broaden the empirical evaluation, they are not necessary to establish the underlying protocol contributions.

Q2. We would like to clarify that requiring only ℓ≤k−1\ell\leq k-1 is an advantage of our protocol rather than a limitation. Existing truncation approaches relying on the statistical security parameter κ\kappa require a substantially larger gap, resulting in larger algebraic structures and increased computational and communication overhead. Our protocol eliminates this dependence while preserving exact truncation. The signed ℓ\ell-bit input range is a standard correctness assumption in fixed-point MPC, rather than a limitation specific to our protocol. In our experiments, we set ℓ=60\ell=60 and k=64k=64, and verified through plaintext fixed-point tests that intermediate values remain within the required range. When this assumption is violated, the correctness guarantee no longer applies, rather than merely introducing bounded approximation errors.

Thanks for the responses. I have changed my ratings accordingly.

Reviewer_mBGWOpenReview ↗
Rating
2
Confidence
3(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies deterministic fixed-point truncation for general n-party Shamir sharing over Galois rings GR(2k,d) in the semi-honest, honest-majority setting. It folds modular wraparound into the truncation relation, obtaining a γ-free protocol that needs only one online secure comparison when ℓ ≤ k−1. It then extends the protocol to RMFE packing and proposes MulRescale for fused multiplication/rescaling and AggRescale for gradient aggregation. Experiments on MNIST with FCN, CNN, and LeNet models report up to a 1.70× LAN speedup over a prior RMFE-based training framework.

Critical strengths
  1. Folding the wraparound indicator s = rmsb(1 − cmsb) into the truncation relation removes the statistical security parameter and needs only one online comparison when ℓ ≤ k − 1. The correctness proof in Appendix A and the UC proof in Appendix C.2 are easy to verify, and the reduction in ring bit-width from 128 to 64 bits is a concrete improvement over Zhou (2026) [1].

  2. The RMFE-native operators handle a genuine packing problem. MulRescale masks a degree-2t intermediate that may lie outside Im(Pack) by combining a packed mask with a full-ring randomizer; Appendix C.4 proves the mask is uniform over GR(2k,d) while preserving component-wise masks. AggRescale cleanly reduces packed gradient aggregation to scalar truncation. The code link and LAN/WAN, 3 to 15 parties experiments add useful engineering value.

[1] Scalable privacy-preserving neural network training over Z2^k via RMFE-based packing and mixed-circuit computation.

Critical weaknesses
  1. The truncation formulation and RMFE operators are clean, but the technical contributions appear to be largely incremental extensions of existing techniques. Eq. (2) is essentially the MSB wraparound correction in Dalskov et al. (2020) [1] plus the standard lower-bit carry comparison and signed offset. Once combined, the one-comparison γ-free protocol follows; no step is identified that cannot be obtained this way. Sec. 4.3 and Appendices C.4–C.5 also closely follow Zhou (2026) [2]'s RMFE framework. The delta is mainly ring width and one comparison, not a new technique.

  2. Training uses only MNIST and small FCN/CNN/LeNet models with 3 parties. WAN speedups are 1.01–1.12×, at 19.95–32.04 h and 36.23–486.94 GB per party per epoch; 15 epochs on Network C would need about 20 days and 7.3 TB per party. No larger dataset, modern model, GPU result, or comparison with other general-n-party frameworks is given. Deployment value is unconvincing.

  3. Table 3 omits the offline/online split and storage. Table 2 pairs DTrun at ρ = 2 with edaBit at ρ = 1, omitting Zhou's ρ = 2 edaBit. The claimed WAN 100 MB/s conflicts with Zhou's 100 Mbps, and the baseline numbers match Zhou's exactly.

  4. Table 3 compares the complete systems; Fig. 1 varies packing, pooling, and loss. Neither separates truncation, fused multiplication, and aggregation. A controlled component-level breakdown would help clarify the individual contributions of these optimizations. It would also be helpful to include the packed edaBit baseline.

  5. The 64-bit configuration depends on every pre-truncation product and aggregate satisfying the signed 60-bit bound (Corollary 3.2; Appendices C.4-C.5). No intermediate-range measurements or overflow checks are reported. Accuracy curves alone do not establish this condition. Report the ranges and check outputs against a plaintext fixed-point reference.

[1] Secure evaluation of quantized neural net-works.

[2] Scalable privacy-preserving neural network training over Z2^k via RMFE-based packing and mixed-circuit computation.

Questions for authors
  1. Which step of DTrun cannot be obtained by combining Dalskov et al. (2020) [1]'s MSB correction with the standard carry comparison and signed offset? Please state the concrete technical obstacle and the additional guarantee over Zhou (2026) [2].

  2. How do you ensure that every degree-2t intermediate and gradient aggregate stays within the signed 60-bit bound? Please report measured ranges or overflow checks and compare secure outputs with a plaintext fixed-point implementation.

  3. Was the WAN bandwidth 100 MB/s or 100 Mbps? Were all Table 2 baselines rerun in your environment? Please separate offline preprocessing from online training, report storage, and add the packed edaBit ρ = 2 baseline.

  4. Which realistic deployment do the MNIST results support, given 20–32 h and 36–487 GB per party per epoch in WAN? How does the system compare with other general-n-party training frameworks on larger models or datasets?

  5. How are the signed range bounds maintained during products, accumulated dot products, and gradient aggregation? Were the secure outputs checked against a plaintext fixed-point implementation?

[1] Secure evaluation of quantized neural net-works.

[2] Scalable privacy-preserving neural network training over Z2^k via RMFE-based packing and mixed-circuit computation.

1 comment

We sincerely thank the reviewer for their valuable time and thoughtful feedback. Below, we address the key concerns raised in the review.

Q1. Relation to TruncPrZ2k and Technical Novelty

Concrete technical differences. While Dalskov et al.'s MSB correction can be combined with a carry comparison and signed offset to recover an algebraically equivalent scalar formula under ℓ≤k−1\ell\leq k-1, several technical differences remain. First, TruncPrZ2k provides arithmetic-only randomness rather than the mixed-circuit edaBits required for BitLT. Second, rather than extending probabilistic truncation with additional corrections, we directly characterize modular wraparound over a more general modulus NN and derive a unified exact truncation relation (Theorem 3.1). This provides an alternative algebraic perspective that is not restricted to power-of-two moduli. Specializing to Z2k\mathbb{Z}_{2^k} yields a κ\kappa-free deterministic protocol with only one online secure comparison, compared with two sequential comparisons in Escudero et al. Third, RMFE-native computation introduces an additional challenge: degree-2t2t multiplication outputs may lie outside Im⁡(Pack)\operatorname{Im}(\mathrm{Pack}). Our ΠMulRescale\Pi_{\mathrm{MulRescale}} addresses this through full-ring masking, combining packed truncation masks with kernel randomness to securely reveal the masked intermediate without separate degree reduction.

Additional guarantees over Zhou (2026). Our truncation protocol achieves perfect security in the stated hybrid model (Theorem 4.1), eliminating the statistical security parameter κ\kappa while retaining exact deterministic floor rounding. Moreover, ΠMulRescale\Pi_{\mathrm{MulRescale}} and ΠAggRescale\Pi_{\mathrm{AggRescale}} extend these guarantees to multiplication-rescaling and gradient aggregation, respectively.

Q2 & Q5. Signed Input Range and Validation.

As is standard in fixed-point MPC protocols, correctness assumes that values entering truncation, including intermediate multiplication results and gradient aggregates, remain within a specified signed ℓ\ell-bit range. This is not a restriction unique to our construction. In our experiments, we set ℓ=60\ell=60 and used plaintext fixed-point tests to verify that these intermediate values remain within the required range for the evaluated configurations. Thus, our experimental settings satisfy this standard assumption, under which our protocols guarantee exact correctness.

Q3.

We thank the reviewer for identifying the WAN bandwidth inconsistency. The correct configuration is 100 Mbps with 40 ms RTT, consistent with Zhou (2026). The reported 100 MB/s is a typographical error, which we will correct. Since the ΠedaBit\Pi_{\mathrm{edaBit}} baseline does not employ packing, we compare it with ΠDTrun\Pi_{\mathrm{DTrun}} under the same non-packed setting (ρ=1\rho=1, k=64k=64) in Table 2. This provides a fair comparison that isolates the efficiency improvements of our truncation protocol from the benefits of RMFE packing. For the RMFE-based baseline of Zhou (2026), which also employs packing, we additionally report results with ρ=2\rho=2 to demonstrate the efficiency gains in the packed setting. At the operator level, Table 2 provides a detailed breakdown of both offline and online costs, allowing us to evaluate the efficiency improvements in each phase. Table 3 further evaluates online end-to-end training performance, focusing on the latency-critical phase of practical training under the offline/online preprocessing model.

Q4.

Our experiments aim to demonstrate efficiency improvements within the existing RMFE-based training framework rather than establish large-scale deployment readiness. Beyond the three-party end-to-end training experiments, we evaluate scalability from 3 to 15 parties, demonstrating favorable scalability as the number of participants increases.

Reviewer_C1JzOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a deterministic truncation protocol for general (n)-party Shamir secret sharing over Galois rings. By integrating modular wraparound correction into the truncation formula, the protocol eliminates the statistical security parameter (\kappa) while requiring only one online secure comparison. The authors further extend the protocol to RMFE-based fixed-point multiplication and gradient aggregation, achieving up to a 1.70× speedup in private neural network training.

Critical strengths

The proposed protocol achieves both (\kappa)-free truncation and a single online secure comparison, improving upon [1], which requires two sequential comparisons, and [2], which depends on (\kappa). Furthermore, the method is particularly suitable for deterministic truncation in general (n)-party MPC, a setting not directly supported by existing 2PC or fixed-party replicated-secret-sharing frameworks such as ABY3 and SecureNN.

[1] Daniel Escudero, Satrajit Ghosh, Marcel Keller, Rahul Rachuri, and Peter Scholl. Improved primitives for MPC over mixed arithmetic-binary circuits. In Advances in Cryptology – CRYPTO 2020, pp. 823–852, 2020.

[2] Hengcheng Zhou. Scalable privacy-preserving neural network training over Z2k via RMFE-based packing and mixed-circuit computation. Proceedings of the AAAI Conference on Artificial Intelligence, 40(34):28964–28972, Mar. 2026.

Critical weaknesses

The novelty and practical significance of the contributions appear somewhat limited. In particular, the MSB-based wraparound detection technique is closely related to Theorem 3.2 of [3], which already identifies wraparound using the MSBs of the random mask and masked value with only a 1-bit gap, albeit in a different algebraic setting. Furthermore, the AI-related contributions largely build upon the existing RMFE-based training framework, with most modifications focusing on eliminating (\kappa). The paper also lacks comparisons with competitive probabilistic-truncation-based training frameworks, leaving its broader practical advantages unclear.

[3] Fengrun Liu, Xie Xiang, and Yu Yu. Scalable multi-party computation protocols for machine learning in the honest-majority setting. In Proceedings of the 33rd USENIX Conference on Security Symposium, 2024.

Questions for authors

Could the authors clarify when deterministic truncation is preferable to probabilistic truncation in private neural network training? For example, [4] employs probabilistic truncation and achieves competitive or even higher accuracy on comparable tasks while supporting up to ten parties. Furthermore, the paper motivates deterministic truncation partly through security vulnerabilities in probabilistic truncation. Are these vulnerabilities inherent to probabilistic truncation or specific to certain constructions? Given subsequent developments, including [5], a more comprehensive discussion and empirical comparison would help justify the practical necessity of the proposed approach.

[4] Marcel Keller and Ke Sun. Secure quantized training for deep learning. In ICML, 2022.

[5] Lijing Zhou, Bingsheng Zhang, Ziyu Wang, Tianpei Lu, Qingrui Song, Su Zhang, Hongrui Cui, and Yu Yu. On probabilistic truncation in privacy-preserving machine learning. Proceedings of the AAAI Conference on Artificial Intelligence, 39(21):22955–22964, Apr. 2025.

1 comment

We sincerely thank the reviewer for their valuable time and thoughtful feedback. Below, we address the key concerns raised in the review.

  1. Deterministic vs. Probabilistic Truncation. Although probabilistic truncation can achieve competitive execution times, its output is a deterministic function of both the input and the masking randomness. As shown by Li et al. (USENIX Security 2023), reusing this randomness for both input masking and rounding can introduce correlations between the output and the revealed transcript, compromising the security of several existing protocols. Therefore, our work focuses exclusively on deterministic truncation, achieving exact correctness and perfect security under the stated security model.
  2. Technical Novelty. Our technical contributions lie in two aspects. First, rather than extending existing truncation protocols with additional corrections, we directly characterize modular wraparound over a broader class of moduli NN and derive a unified exact truncation relation. This provides an alternative algebraic perspective that is not restricted to power-of-two moduli. When specialized to Z2k\mathbb{Z}_{2^k}, the lower-bit carry correction becomes independent of wraparound, enabling κ\kappa-free deterministic truncation with only one online secure comparison. Second, we develop two RMFE-native operators, ΠMulRescale\Pi_{\mathrm{MulRescale}} and ΠAggRescale\Pi_{\mathrm{AggRescale}}, for multiplication-rescaling and gradient aggregation, respectively. In particular, ΠMulRescale\Pi_{\mathrm{MulRescale}} introduces a full-ring masking construction that uniformly masks arbitrary Galois-ring intermediates while preserving component-wise truncation masks. This addresses the challenge that degree-2t2t multiplication results may lie outside Im⁡(Pack)\operatorname{Im}(\mathrm{Pack}), enabling direct truncation into valid degree-tt packed sharings without a separate degree-reduction step.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.