Efficient and High-Accuracy Private CNN Inference with Helper-Assisted Malicious Security

Junchao Fan; Kaiwen Wang; Xiaolin Chang; Yuehan Dong

arxiv: 2507.09607 · v4 · submitted 2025-07-13 · 💻 cs.CR

Efficient and High-Accuracy Private CNN Inference with Helper-Assisted Malicious Security

Kaiwen Wang , Xiaolin Chang , Junchao Fan , Yuehan Dong This is my paper

Pith reviewed 2026-05-19 05:13 UTC · model grok-4.3

classification 💻 cs.CR

keywords private CNN inferencemalicious securitysecure computationpolynomial activationknowledge distillationfixed-point arithmeticMPC protocolshelper-assisted model

0 comments

The pith

Co-designed ring sharing and constant-round protocols enable private CNN inference under malicious security to run several times faster while staying within 0.5% of ReLU accuracy.

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

The paper works to resolve the tension among strong malicious security, low communication cost, and high model accuracy in private convolutional neural network inference. It does so by extending authenticated secret sharing to rings for efficient fixed-point arithmetic, creating multiplication and polynomial-evaluation protocols whose round count does not grow with polynomial degree, and training the resulting polynomial networks with knowledge distillation and warm initialization. A sympathetic reader would care because the approach removes the need to choose between security against malicious servers and practical speed or accuracy when clients query models on private data. If the claims hold, service providers could deliver accurate CNN results on sensitive inputs without learning those inputs, even when the server may deviate arbitrarily from the protocol.

Core claim

By extending authenticated sharing to rings, designing constant-round protocols for multiplication and polynomial evaluation independent of degree, and applying knowledge distillation with warm initialization during training, the helper-assisted malicious dishonest majority framework for private CNN inference achieves 2.3 to 6.8 times speedup in LAN and 1.3 to 5.6 times in WAN over prior state-of-the-art malicious secure methods while keeping accuracy within 0.5 percent of ReLU-based plaintext models.

What carries the argument

Ring-extended authenticated secret sharing together with constant-round, degree-independent protocols for multiplication and polynomial evaluation, paired with knowledge-distillation training.

Load-bearing premise

The knowledge-distillation and warm-initialization training recovers enough expressiveness from polynomial activations to reach accuracy within 0.5 percent of ReLU models on the evaluated CNNs.

What would settle it

Running the same CNN architectures on the same datasets with the proposed training and observing accuracy more than 0.5 percent below the ReLU baseline would falsify the accuracy claim.

read the original abstract

Machine Learning as a Service (MLaaS) exposes sensitive client data to service providers. Private inference mitigates this risk while preserving model functionality. Despite extensive progress in MPC-based solutions, they remain constrained by a fundamental three-way tension among strong security, efficiency, and model accuracy. This challenge is particularly acute under the malicious dishonest majority (MSDM) setting, where prior work either incurs high communication overhead or suffers non-negligible accuracy loss due to polynomial approximations of nonlinear functions. Although the helper-assisted MSDM (HA-MSDM) model improves efficiency and fairness, it lacks a dedicated design for accurate and efficient neural network inference. In this work, we present an HA-MSDM-based private CNN inference framework that simultaneously achieves high efficiency and near-plaintext accuracy through a co-design of cryptographic primitives, MPC protocols, and model training. Specifically, we (i) extend authenticated sharing to rings to enable efficient fixed-point computation, (ii) design constant-round protocols for multiplication and polynomial evaluation, with round complexity independent of the polynomial degree, and (iii) introduce a training strategy that recovers the expressiveness of polynomial models via knowledge distillation and warm initialization. Experiments demonstrate 2.3--6.8$\times$ speedup in LAN and 1.3--5.6$\times$ in WAN over state-of-the-art MSDM frameworks, while achieving accuracy within 0.5\% of ReLU-based plaintext models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper delivers a co-designed HA-MSDM framework with ring shares, constant-round polynomial protocols, and distillation training that produces real speedups while keeping accuracy close to ReLU baselines.

read the letter

This paper's main advance is a practical HA-MSDM private CNN inference system that combines ring-based authenticated sharing for fixed-point arithmetic, constant-round protocols for multiplication and polynomial evaluation, and a knowledge-distillation plus warm-initialization training method. The protocols look like a genuine efficiency step over earlier MSDM work because the round count stays independent of polynomial degree, and the ring extension fits the fixed-point needs without extra overhead. The training co-design is a reasonable way to restore model power without pushing to high-degree polynomials that would hurt performance. Experiments report 2.3-6.8× LAN and 1.3-5.6× WAN speedups with accuracy within 0.5% of plaintext ReLU models, which is the kind of concrete gain that matters for deployment questions. The constructions appear formally grounded and the citation pattern to prior MSDM and HA-MSDM papers is appropriate. The central accuracy claim rests on the training strategy actually closing the usual polynomial gap for the chosen CNN depths and datasets. If the paper includes ablations that isolate the distillation contribution versus architecture or degree choices, that would make the result more robust; without them the 0.5% figure could be narrower than stated. The experimental section would benefit from error bars, exact baseline implementations, and dataset details to let readers judge the controls. This is useful reading for researchers already working on malicious-secure MPC for ML or on practical private inference. A reader focused on protocol efficiency or on closing the accuracy gap in helper-assisted settings will find the concrete numbers and protocol descriptions worth their time. It deserves a serious referee because the protocol ideas are new enough and the problem is relevant, even if the experiments need some tightening.

Referee Report

2 major / 2 minor

Summary. The paper proposes an HA-MSDM private CNN inference framework that co-designs cryptographic primitives (authenticated sharing over rings for fixed-point arithmetic), MPC protocols (constant-round multiplication and polynomial evaluation whose round complexity is independent of degree), and model training (knowledge distillation plus warm initialization) to achieve both high efficiency under malicious dishonest-majority security and accuracy within 0.5% of ReLU-based plaintext models. Experiments are reported to show 2.3–6.8× LAN and 1.3–5.6× WAN speedups over prior MSDM frameworks.

Significance. If the experimental claims are reproducible, the work would meaningfully advance practical private inference by simultaneously improving efficiency and closing the accuracy gap that typically arises from polynomial activation approximations under strong malicious security. The constant-round polynomial-evaluation protocol is a technically interesting primitive that could apply beyond this setting.

major comments (2)

[§5] §5 (Evaluation) and abstract: the headline accuracy claim of 'within 0.5% of ReLU-based plaintext models' rests on the unverified effectiveness of the knowledge-distillation plus warm-initialization strategy. No ablation studies, error bars, exact dataset splits, or comparisons against standard training on the same polynomial degrees and CNN architectures are provided, so it is impossible to isolate whether the reported accuracy is attributable to the co-design or to the specific models and datasets chosen.
[§5] §5 and Table 2 (or equivalent performance table): the reported speedups lack details on exact baseline implementations, network depths, polynomial degrees used in each comparison, and whether the ReLU baselines were trained identically to the polynomial models. Without these controls the 2.3–6.8× and 1.3–5.6× figures cannot be confidently attributed to the new protocols alone.

minor comments (2)

[§3] Notation for ring elements and authenticated shares in §3 should be made fully consistent with the fixed-point representation used in the protocols.
[§2] The security definition and threat model for the helper-assisted setting could be stated more explicitly in §2 to clarify how the helper affects the malicious dishonest-majority assumption.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and will revise the evaluation section to incorporate the requested clarifications and additional experimental details. These revisions will strengthen the reproducibility and attribution of our results while preserving the core technical contributions of the HA-MSDM framework.

read point-by-point responses

Referee: [§5] §5 (Evaluation) and abstract: the headline accuracy claim of 'within 0.5% of ReLU-based plaintext models' rests on the unverified effectiveness of the knowledge-distillation plus warm-initialization strategy. No ablation studies, error bars, exact dataset splits, or comparisons against standard training on the same polynomial degrees and CNN architectures are provided, so it is impossible to isolate whether the reported accuracy is attributable to the co-design or to the specific models and datasets chosen.

Authors: We appreciate the referee's emphasis on isolating the contribution of our training strategy. Section 4.3 of the manuscript describes the knowledge-distillation procedure and warm-initialization technique in detail, and Section 5 reports the final top-1 accuracies achieved on the evaluated CNNs. However, we acknowledge that explicit ablation tables (with/without distillation), standard deviations from multiple independent runs, precise train/validation/test splits, and direct comparisons to standard training on identical polynomial degrees and architectures are not present. In the revised manuscript we will add these ablations and statistical details to Section 5, enabling readers to attribute accuracy recovery specifically to the proposed co-design rather than to dataset or architecture idiosyncrasies. revision: yes
Referee: [§5] §5 and Table 2 (or equivalent performance table): the reported speedups lack details on exact baseline implementations, network depths, polynomial degrees used in each comparison, and whether the ReLU baselines were trained identically to the polynomial models. Without these controls the 2.3–6.8× and 1.3–5.6× figures cannot be confidently attributed to the new protocols alone.

Authors: We agree that precise baseline specifications are essential for attributing the observed speedups. The manuscript already states the network architectures (e.g., ResNet-18/34 variants) and the polynomial degrees employed in our models, and compares against the most recent published MSDM implementations. Nevertheless, we recognize that exact baseline code references, layer-wise network depths, the precise polynomial degrees used by each prior work, and confirmation that ReLU baselines received identical training hyperparameters are not fully enumerated in Table 2 or the surrounding text. In the revision we will expand Section 5 and Table 2 with these controls, including footnotes that list the original papers' reported polynomial degrees and training protocols, thereby clarifying that the speedups derive from our constant-round multiplication and polynomial-evaluation protocols rather than from differences in model configuration. revision: yes

Circularity Check

0 steps flagged

No circularity: new protocols and empirical training co-design are self-contained

full rationale

The paper constructs new authenticated sharing extensions, constant-round multiplication and polynomial-evaluation protocols, and a knowledge-distillation plus warm-initialization training procedure. Reported speedups (2.3–6.8× LAN, 1.3–5.6× WAN) and accuracy (within 0.5 % of ReLU plaintext) are presented as experimental outcomes on specific CNNs and datasets, not as quantities derived from or fitted to the same inputs by construction. No equations, uniqueness theorems, or self-citations are shown that would reduce the central claims to tautological re-labeling of prior fitted values or self-referential definitions. The derivation chain therefore remains independent of its own outputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The framework rests on the HA-MSDM security model and standard cryptographic assumptions for authenticated sharing and MPC; no new free parameters, invented entities, or ad-hoc axioms are introduced in the abstract description.

axioms (1)

domain assumption The helper-assisted malicious dishonest majority model provides the stated security and efficiency properties when the helper follows the protocol.
The entire framework is built on top of the HA-MSDM setting introduced in prior work.

pith-pipeline@v0.9.0 · 5794 in / 1291 out tokens · 35535 ms · 2026-05-19T05:13:23.583763+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We propose six efficient MPC protocols... multiplication, multiplication truncation, exponentiation... two-party polynomial protocols... sixth-order polynomial approximation... parameter-adjusted batch normalization layer
Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Experiments demonstrate 2.3--6.8× speedup... accuracy within 0.5% of ReLU-based plaintext models

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

37 extracted references · 37 canonical work pages

[1]

https://www.solutelabs.com/blog/best-mlaas-platforms, 2025

work page 2025
[2]

Lucien K. L. Ng, Sherman S. M. Chow: SoK: Cryptographic Neural - Network Computation. SP 2023: 497-514

work page 2023
[3]

FOCS 1982: 160-164

Andrew Chi -Chih Yao: Protocols for Secure Computations (Extended Abstract). FOCS 1982: 160-164

work page 1982
[4]

CRYPTO (2) 2018: 769-798

Ronald Cramer, Ivan Damgå rd, Daniel Escudero, Peter Scholl, Chaoping Xing:SPDℤ2k: Efficient MPC mod 2k for Dishonest Majority. CRYPTO (2) 2018: 769-798

work page 2018
[5]

USENIX Security Symposium 2024

Boshi Yuan, Shixuan Yang, Yongxiang Zhang, Ning Ding, Dawu Gu, Shi-Feng Sun: MD-ML: Super Fast Privacy -Preserving Machine Learning for Malicious Security with a Dishonest Majority. USENIX Security Symposium 2024

work page 2024
[6]

IEEE Trans

Yansong Zhang, Xiaojun Chen, Ye Dong, Qinghui Zhang, Rui Hou, Qiang Liu, Xudong Chen: MD-SONIC: Maliciously-Secure Outsourcing Neural Network Inference With Reduced Online Communication. IEEE Trans. Inf. Forensics Secur. 20: 3534-3549 (2025)

work page 2025
[7]

Wahby, Fraser Brown, Wenting Zheng: Silph: A Framework for Scalable and Accurate Generation of Hybrid MPC Protocols

Edward Chen, Jinhao Zhu, Alex Ozdemir, Riad S. Wahby, Fraser Brown, Wenting Zheng: Silph: A Framework for Scalable and Accurate Generation of Hybrid MPC Protocols. SP 2023: 848-863

work page 2023
[8]

USENIX Security Symposium 2022: 827-844

Jean-Luc Watson, Sameer Wagh, Raluca Ada Popa: Piranha: A GPU Platform for Secure Computation. USENIX Security Symposium 2022: 827-844

work page 2022
[9]

USENIX Security Symposium 2024

Fengrun Liu, Xiang Xie, Yu: Scalable Multi-Party Computation Protocols for Machine Learning in the Honest -Majority Setting. USENIX Security Symposium 2024

work page 2024
[10]

USENIX Security Symposium 2021: 2651-2668

Nishat Koti, Mahak Pancholi, Arpita Patra, Ajith Suresh: SWIFT: Super- fast and Robust Privacy-Preserving Machine Learning. USENIX Security Symposium 2021: 2651-2668

work page 2021
[11]

Sameer Wagh, Shruti Tople, Fabrice Benhamouda, Eyal Kushilevitz, Prateek Mittal, Tal Rabin: Falcon: Honest-Majority Maliciously Secure Framework for Private Deep Learning. Proc. Priv. Enhancing Technol. 2021(1): 188-208 (2021)

work page 2021
[12]

SP 2024: 542- 560

Banashri Karmakar, Nishat Koti, Arpita Patra, Sikhar Patranabis, Protik Paul, Divya Ravi: Asterisk: Super-fast MPC with a Friend. SP 2024: 542- 560

work page 2024
[13]

STOC 1986: 364-369

Richard Cleve:Limits on the Security of Coin Flips when Half the Processors Are Faulty (Extended Abstract). STOC 1986: 364-369

work page 1986
[14]

SP 2024: 4753-4771

Qi Pang, Jinhao Zhu, Helen Mö llering, Wenting Zheng, Thomas Schneider: BOLT: Privacy-Preserving, Accurate and Efficient Inference for Transformers. SP 2024: 4753-4771

work page 2024
[15]

CCS 2021: 3266-3281

Siam Umar Hussain, Mojan Javaheripi, Mohammad Samragh, Farinaz Koushanfar: COINN: Crypto/ML Codesign for Oblivious Inference via Neural Networks. CCS 2021: 3266-3281

work page 2021
[16]

NeurIPS 2020

Zahra Ghodsi, Akshaj Kumar Veldanda, Brandon Reagen, Siddharth Garg: CryptoNAS: Private Inference on a ReLU Budget. NeurIPS 2020

work page 2020
[17]

CoRR abs/2107.12342 (2021)

Karthik Garimella, Nandan Kumar Jha, Brando n Reagen: Sisyphus: A Cautionary Tale of Using Low -Degree Polynomial Activations in Privacy-Preserving Deep Learning. CoRR abs/2107.12342 (2021)

work page arXiv 2021
[18]

USENIX Security Symposium 2024

Abdulrahman Diaa, Lucas Fenaux, Thomas Humphries, Marian Dietz, Faezeh Ebrahimianghazani, Bailey Kacsmar, Xinda Li, Nils Lukas, Rasoul Akhavan Mahdavi, Simon Oya, Ehsan Amjadian, Florian Kerschbaum: Fast and Private Inference of Deep Neural Networks by Co- designing Activation Functions. USENIX Security Symposium 2024

work page 2024
[19]

CCS 2019: 887-903

Donghang Lu, Thomas Yurek, Samarth Kulshreshtha, Rahul Govind, Aniket Kate, Andrew Miller: HoneyBadgerMPC and AsynchroMix: Practical Asynchronous MPC and its Application to Anonymous Communication. CCS 2019: 887-903

work page 2019
[20]

IEEE Symposium on Security and Privacy 2017: 19-3

Payman Mohassel, Yupeng Zhang: SecureML: A System for Scalable Privacy-Preserving Machine Learning. IEEE Symposium on Security and Privacy 2017: 19-3

work page 2017
[21]

Smart, Sarah Zakarias: Multiparty Computation from Somewhat Homomorphic Encryption

Ivan Damgå rd, Valerio Pastro, Nigel P. Smart, Sarah Zakarias: Multiparty Computation from Somewhat Homomorphic Encryption. CRYPTO 2012: 643-662

work page 2012
[22]

Smart: Practical Covertly Secure MPC for Dishonest Majority - Or: Breaking the SPDZ Limits

Ivan Damgå rd, Marcel Keller, Enrique Larraia, Valerio Pastro, Peter Scholl, Nigel P. Smart: Practical Covertly Secure MPC for Dishonest Majority - Or: Breaking the SPDZ Limits. ESORICS 2013: 1-188

work page 2013
[23]

CCS 2016: 830-8422

Marcel Keller, Emmanuela Orsini, Peter Scholl: MASCOT: Faster Malicious Arithmetic Secure Computation with Oblivious Transfer. CCS 2016: 830-8422

work page 2016
[24]

EUROCRYPT (3) 2018: 158-189

Marcel Keller, Valerio Pastro, Dragos Rotaru: Overdrive: Making SPDZ Great Again. EUROCRYPT (3) 2018: 158-189

work page 2018
[25]

Smart, Frederik Vercauteren: Overdrive2k: Efficient Secure MPC over $ \mathbb {Z}_{2k}$

Emmanuela Orsini, Nigel P. Smart, Frederik Vercauteren: Overdrive2k: Efficient Secure MPC over $ \mathbb {Z}_{2k}$ . from Somewhat Homomorphic Encryption. CT-RSA 2020: 254-283

work page 2020
[26]

CRYPTO (1) 2022: 383-412

Daniel Escudero, Chaoping Xing, Chen Yuan: More Efficient Dishonest Majority Secure Computation over $\mathbb {Z}_{2k}$ via Galois Rings. CRYPTO (1) 2022: 383-412

work page 2022
[27]

CRYPTO (3) 2019: 432-461

Nishanth Chandran, Wutichai Chongchitmate, Rafail Ostrovsky, Ivan Visconti: Universally Composable Secure Computation with Corrupted Tokens. CRYPTO (3) 2019: 432-461

work page 2019
[28]

TCC (B1) 2016: 367-399

Carmit Hazay, Antigoni Polychroniadou, Muthuramakrishnan Venkitasubramaniam: Composable Security in the Tamper -Proof Hardware Model Under Minimal Complexity. TCC (B1) 2016: 367-399

work page 2016
[29]

ASIACRYPT (3) 2018: 118-138

Saikrishna Badrinarayanan, Abhishek Jain, Rafail Ostrovsky, Ivan Visconti: Non-interactive Secure Computation from One-Way Functions. ASIACRYPT (3) 2018: 118-138

work page 2018
[30]

TCC (2) 2022: 470 - 501

Yuval Ishai, Arpita Patra, Sikhar Patranabis, Divya Ravi, Akshayaram Srinivasan: Fully-Secure MPC with Minimal Trust. TCC (2) 2022: 470 - 501

work page 2022
[31]

Cryptology ePrint Archive, 2022

Philipp Muth, and Stefan Katzenbeisser: Assisted mpc. Cryptology ePrint Archive, 2022

work page 2022
[32]

IEEE Symposium on Security and Privacy 2019: 1102-1120

Ivan Damgå rd, Daniel Escudero, Tore Kasper Frederiksen, Marcel Keller, Peter Scholl, Nikolaj Volgushev: New Primitives for Actively -Secure MPC over Rings with Applications to Private Machine Learning. IEEE Symposium on Security and Privacy 2019: 1102-1120

work page 2019
[33]

Lauter, Michael Naehrig, John Wernsing: CryptoNets: Applying Neural Networks to Encrypted Data with High Throughput and Accuracy

Ran Gilad -Bachrach, Nathan Dowlin, Kim Laine, Kristin E. Lauter, Michael Naehrig, John Wernsing: CryptoNets: Applying Neural Networks to Encrypted Data with High Throughput and Accuracy. ICML 2016: 201-210

work page 2016
[34]

IACR Cryptol

Hervé Chabanne, Amaury de Wargny, Jonathan Milgram, Constance Morel, Emmanuel Prouff: Privacy-Preserving Classification on Deep Neural Network. IACR Cryptol. ePrint Arch. 2017: 35 (2017)

work page 2017
[35]

Wright: Privacy-preserving Machine Learning as a Service

Ehsan Hesamifard, Hassan Takabi, Mehdi Ghasemi, Rebecca N. Wright: Privacy-preserving Machine Learning as a Service. Proc. Priv. Enhancing Technol. 2018(3): 123-142 (2018)

work page 2018
[36]

USENIX Security Symposium 2020: 2505-2522

Pratyush Mishra, Ryan Lehmkuhl, Akshayaram Srinivasan, Wenting Zheng, Raluca Ada Popa: Delphi: A Cryptographic Inference Service for Neural Networks. USENIX Security Symposium 2020: 2505-2522

work page 2020
[37]

IEEE Access 11: 62062- 62076 (2023)

Junghyun Lee, Eunsang Lee, Joon-Woo Lee, Yongjune Kim, Young-Sik Kim, Jong -Seon No: Precise Approximation of Convolutional Neural Networks for Homomorphically Encrypted Data. IEEE Access 11: 62062- 62076 (2023)

work page 2023

[1] [1]

https://www.solutelabs.com/blog/best-mlaas-platforms, 2025

work page 2025

[2] [2]

Lucien K. L. Ng, Sherman S. M. Chow: SoK: Cryptographic Neural - Network Computation. SP 2023: 497-514

work page 2023

[3] [3]

FOCS 1982: 160-164

Andrew Chi -Chih Yao: Protocols for Secure Computations (Extended Abstract). FOCS 1982: 160-164

work page 1982

[4] [4]

CRYPTO (2) 2018: 769-798

Ronald Cramer, Ivan Damgå rd, Daniel Escudero, Peter Scholl, Chaoping Xing:SPDℤ2k: Efficient MPC mod 2k for Dishonest Majority. CRYPTO (2) 2018: 769-798

work page 2018

[5] [5]

USENIX Security Symposium 2024

Boshi Yuan, Shixuan Yang, Yongxiang Zhang, Ning Ding, Dawu Gu, Shi-Feng Sun: MD-ML: Super Fast Privacy -Preserving Machine Learning for Malicious Security with a Dishonest Majority. USENIX Security Symposium 2024

work page 2024

[6] [6]

IEEE Trans

Yansong Zhang, Xiaojun Chen, Ye Dong, Qinghui Zhang, Rui Hou, Qiang Liu, Xudong Chen: MD-SONIC: Maliciously-Secure Outsourcing Neural Network Inference With Reduced Online Communication. IEEE Trans. Inf. Forensics Secur. 20: 3534-3549 (2025)

work page 2025

[7] [7]

Wahby, Fraser Brown, Wenting Zheng: Silph: A Framework for Scalable and Accurate Generation of Hybrid MPC Protocols

Edward Chen, Jinhao Zhu, Alex Ozdemir, Riad S. Wahby, Fraser Brown, Wenting Zheng: Silph: A Framework for Scalable and Accurate Generation of Hybrid MPC Protocols. SP 2023: 848-863

work page 2023

[8] [8]

USENIX Security Symposium 2022: 827-844

Jean-Luc Watson, Sameer Wagh, Raluca Ada Popa: Piranha: A GPU Platform for Secure Computation. USENIX Security Symposium 2022: 827-844

work page 2022

[9] [9]

USENIX Security Symposium 2024

Fengrun Liu, Xiang Xie, Yu: Scalable Multi-Party Computation Protocols for Machine Learning in the Honest -Majority Setting. USENIX Security Symposium 2024

work page 2024

[10] [10]

USENIX Security Symposium 2021: 2651-2668

Nishat Koti, Mahak Pancholi, Arpita Patra, Ajith Suresh: SWIFT: Super- fast and Robust Privacy-Preserving Machine Learning. USENIX Security Symposium 2021: 2651-2668

work page 2021

[11] [11]

Sameer Wagh, Shruti Tople, Fabrice Benhamouda, Eyal Kushilevitz, Prateek Mittal, Tal Rabin: Falcon: Honest-Majority Maliciously Secure Framework for Private Deep Learning. Proc. Priv. Enhancing Technol. 2021(1): 188-208 (2021)

work page 2021

[12] [12]

SP 2024: 542- 560

Banashri Karmakar, Nishat Koti, Arpita Patra, Sikhar Patranabis, Protik Paul, Divya Ravi: Asterisk: Super-fast MPC with a Friend. SP 2024: 542- 560

work page 2024

[13] [13]

STOC 1986: 364-369

Richard Cleve:Limits on the Security of Coin Flips when Half the Processors Are Faulty (Extended Abstract). STOC 1986: 364-369

work page 1986

[14] [14]

SP 2024: 4753-4771

Qi Pang, Jinhao Zhu, Helen Mö llering, Wenting Zheng, Thomas Schneider: BOLT: Privacy-Preserving, Accurate and Efficient Inference for Transformers. SP 2024: 4753-4771

work page 2024

[15] [15]

CCS 2021: 3266-3281

Siam Umar Hussain, Mojan Javaheripi, Mohammad Samragh, Farinaz Koushanfar: COINN: Crypto/ML Codesign for Oblivious Inference via Neural Networks. CCS 2021: 3266-3281

work page 2021

[16] [16]

NeurIPS 2020

Zahra Ghodsi, Akshaj Kumar Veldanda, Brandon Reagen, Siddharth Garg: CryptoNAS: Private Inference on a ReLU Budget. NeurIPS 2020

work page 2020

[17] [17]

CoRR abs/2107.12342 (2021)

Karthik Garimella, Nandan Kumar Jha, Brando n Reagen: Sisyphus: A Cautionary Tale of Using Low -Degree Polynomial Activations in Privacy-Preserving Deep Learning. CoRR abs/2107.12342 (2021)

work page arXiv 2021

[18] [18]

USENIX Security Symposium 2024

Abdulrahman Diaa, Lucas Fenaux, Thomas Humphries, Marian Dietz, Faezeh Ebrahimianghazani, Bailey Kacsmar, Xinda Li, Nils Lukas, Rasoul Akhavan Mahdavi, Simon Oya, Ehsan Amjadian, Florian Kerschbaum: Fast and Private Inference of Deep Neural Networks by Co- designing Activation Functions. USENIX Security Symposium 2024

work page 2024

[19] [19]

CCS 2019: 887-903

Donghang Lu, Thomas Yurek, Samarth Kulshreshtha, Rahul Govind, Aniket Kate, Andrew Miller: HoneyBadgerMPC and AsynchroMix: Practical Asynchronous MPC and its Application to Anonymous Communication. CCS 2019: 887-903

work page 2019

[20] [20]

IEEE Symposium on Security and Privacy 2017: 19-3

Payman Mohassel, Yupeng Zhang: SecureML: A System for Scalable Privacy-Preserving Machine Learning. IEEE Symposium on Security and Privacy 2017: 19-3

work page 2017

[21] [21]

Smart, Sarah Zakarias: Multiparty Computation from Somewhat Homomorphic Encryption

Ivan Damgå rd, Valerio Pastro, Nigel P. Smart, Sarah Zakarias: Multiparty Computation from Somewhat Homomorphic Encryption. CRYPTO 2012: 643-662

work page 2012

[22] [22]

Smart: Practical Covertly Secure MPC for Dishonest Majority - Or: Breaking the SPDZ Limits

Ivan Damgå rd, Marcel Keller, Enrique Larraia, Valerio Pastro, Peter Scholl, Nigel P. Smart: Practical Covertly Secure MPC for Dishonest Majority - Or: Breaking the SPDZ Limits. ESORICS 2013: 1-188

work page 2013

[23] [23]

CCS 2016: 830-8422

Marcel Keller, Emmanuela Orsini, Peter Scholl: MASCOT: Faster Malicious Arithmetic Secure Computation with Oblivious Transfer. CCS 2016: 830-8422

work page 2016

[24] [24]

EUROCRYPT (3) 2018: 158-189

Marcel Keller, Valerio Pastro, Dragos Rotaru: Overdrive: Making SPDZ Great Again. EUROCRYPT (3) 2018: 158-189

work page 2018

[25] [25]

Smart, Frederik Vercauteren: Overdrive2k: Efficient Secure MPC over $ \mathbb {Z}_{2k}$

Emmanuela Orsini, Nigel P. Smart, Frederik Vercauteren: Overdrive2k: Efficient Secure MPC over $ \mathbb {Z}_{2k}$ . from Somewhat Homomorphic Encryption. CT-RSA 2020: 254-283

work page 2020

[26] [26]

CRYPTO (1) 2022: 383-412

Daniel Escudero, Chaoping Xing, Chen Yuan: More Efficient Dishonest Majority Secure Computation over $\mathbb {Z}_{2k}$ via Galois Rings. CRYPTO (1) 2022: 383-412

work page 2022

[27] [27]

CRYPTO (3) 2019: 432-461

Nishanth Chandran, Wutichai Chongchitmate, Rafail Ostrovsky, Ivan Visconti: Universally Composable Secure Computation with Corrupted Tokens. CRYPTO (3) 2019: 432-461

work page 2019

[28] [28]

TCC (B1) 2016: 367-399

Carmit Hazay, Antigoni Polychroniadou, Muthuramakrishnan Venkitasubramaniam: Composable Security in the Tamper -Proof Hardware Model Under Minimal Complexity. TCC (B1) 2016: 367-399

work page 2016

[29] [29]

ASIACRYPT (3) 2018: 118-138

Saikrishna Badrinarayanan, Abhishek Jain, Rafail Ostrovsky, Ivan Visconti: Non-interactive Secure Computation from One-Way Functions. ASIACRYPT (3) 2018: 118-138

work page 2018

[30] [30]

TCC (2) 2022: 470 - 501

Yuval Ishai, Arpita Patra, Sikhar Patranabis, Divya Ravi, Akshayaram Srinivasan: Fully-Secure MPC with Minimal Trust. TCC (2) 2022: 470 - 501

work page 2022

[31] [31]

Cryptology ePrint Archive, 2022

Philipp Muth, and Stefan Katzenbeisser: Assisted mpc. Cryptology ePrint Archive, 2022

work page 2022

[32] [32]

IEEE Symposium on Security and Privacy 2019: 1102-1120

Ivan Damgå rd, Daniel Escudero, Tore Kasper Frederiksen, Marcel Keller, Peter Scholl, Nikolaj Volgushev: New Primitives for Actively -Secure MPC over Rings with Applications to Private Machine Learning. IEEE Symposium on Security and Privacy 2019: 1102-1120

work page 2019

[33] [33]

Lauter, Michael Naehrig, John Wernsing: CryptoNets: Applying Neural Networks to Encrypted Data with High Throughput and Accuracy

Ran Gilad -Bachrach, Nathan Dowlin, Kim Laine, Kristin E. Lauter, Michael Naehrig, John Wernsing: CryptoNets: Applying Neural Networks to Encrypted Data with High Throughput and Accuracy. ICML 2016: 201-210

work page 2016

[34] [34]

IACR Cryptol

Hervé Chabanne, Amaury de Wargny, Jonathan Milgram, Constance Morel, Emmanuel Prouff: Privacy-Preserving Classification on Deep Neural Network. IACR Cryptol. ePrint Arch. 2017: 35 (2017)

work page 2017

[35] [35]

Wright: Privacy-preserving Machine Learning as a Service

Ehsan Hesamifard, Hassan Takabi, Mehdi Ghasemi, Rebecca N. Wright: Privacy-preserving Machine Learning as a Service. Proc. Priv. Enhancing Technol. 2018(3): 123-142 (2018)

work page 2018

[36] [36]

USENIX Security Symposium 2020: 2505-2522

Pratyush Mishra, Ryan Lehmkuhl, Akshayaram Srinivasan, Wenting Zheng, Raluca Ada Popa: Delphi: A Cryptographic Inference Service for Neural Networks. USENIX Security Symposium 2020: 2505-2522

work page 2020

[37] [37]

IEEE Access 11: 62062- 62076 (2023)

Junghyun Lee, Eunsang Lee, Joon-Woo Lee, Yongjune Kim, Young-Sik Kim, Jong -Seon No: Precise Approximation of Convolutional Neural Networks for Homomorphically Encrypted Data. IEEE Access 11: 62062- 62076 (2023)

work page 2023