Skip to content

ABY3: A Mixed Protocol Framework for Machine Learning

Authors: Mohassel, Rindal
Year: 2018
Venue: CCS 2018

type: scheme collection: schemes status: auto

Links: ePrint 2018/403

Visual lesson: Replicated secret sharing (3PC, ABY3) · Scheme table row: Replicated secret sharing (3PC, ABY3)

Summary

Auto-extracted summary (unreviewed)

ABY3 is a three-server secure computation framework for privacy-preserving training and inference of linear regression, logistic regression and neural networks, where data owners secret-share their data among three servers. Its main contribution is a complete set of efficient conversions between arithmetic, binary and Yao three-party sharings, plus new fixed-point multiplication techniques for shared decimal values and customised protocols for piecewise polynomial functions, each with semi-honest and malicious-secure variants. The C++ implementation (semi-honest setting) runs 100x to 10000x faster than the best prior work while also reducing communication.

Abstract

Machine learning is widely used to produce models for a range of applications and is increasingly offered as a service by major technology companies. However, the required massive data collection raises privacy concerns during both training and prediction stages. In this paper, we design and implement a general framework for privacy-preserving machine learning and use it to obtain new solutions for training linear regression, logistic regression and neural network models. Our protocols are in a three-server model wherein data owners secret share their data among three servers who train and evaluate models on the joint data using three-party computation (3PC). Our main contribution is a new and complete framework (ABY3) for efficiently switching back and forth between arithmetic, binary, and Yao 3PC which is of independent interest. Many of the conversions are based on new techniques that are designed and optimized for the first time in this paper. We also propose new techniques for fixed-point multiplication of shared decimal values that extends beyond the three-party case, and customized protocols for evaluating piecewise polynomial functions. We design variants of each building block that is secure against malicious adversaries who deviate arbitrarily. We implement our system in C++. Our protocols are up to four orders of magnitude faster than the best prior work, hence significantly reducing the gap between privacy-preserving and plaintext training.
Source: pdf

Extracted details

Field Value Source Confidence
contribution ABY3 framework for switching between arithmetic, binary and Yao 3PC sharings llm 0.85
contribution new fixed-point multiplication techniques for shared decimal values (beyond three parties) llm 0.85
contribution customised protocols for piecewise polynomial functions llm 0.85
contribution malicious-secure variants of each building block llm 0.85
contribution C++ implementation for linear regression, logistic regression and neural network training and inference llm 0.85
schemes secret sharing llm 0.85
schemes garbled circuits llm 0.85
libraries libOTe llm 0.85
models linear regression llm 0.85
models logistic regression llm 0.85
models neural network llm 0.85
tasks privacy-preserving training llm 0.85
tasks privacy-preserving inference llm 0.85
security level computational kappa = 128, statistical lambda = 40 llm 0.85
threat model three-server honest-majority; semi-honest implementation with malicious-secure protocol variants llm 0.85
interaction pure MPC llm 0.85
limitations implementation and benchmarks cover only the semi-honest setting llm 0.85
limitations MPC does not prevent leakage from the trained model itself or from interactions with it llm 0.85
llm relevance medium llm 0.85
key result speed-up over best prior work: 100x to 10000x (semi-honest 3PC, depending on model and problem size) llm 0.85

Scheme card: Replicated secret sharing (3PC, ABY3)

Group 9. Non-HE cryptographic alternatives to CKKS
Kind Primitive
Hardness assumption Honest majority (information-theoretic)
Plaintext type Fixed point / bits
SIMD packing Vectorized
Bootstrapping n/a
Post-quantum Yes
Status Widely used
Fit for LLM inference Used by PUMA for LLaMA-7B in minutes. Requires three non-colluding servers.
Libraries SecretFlow-SPU, MP-SPDZ

Keyword mentions

Counts of vocabulary terms in the full text (first page of occurrence in brackets).

  • scheme: Secret sharing 47 [p.1], Garbled circuits 28 [p.2], MPC 16 [p.2], Oblivious transfer 8 [p.3]
  • library: ABY / ABY3 16 [p.1], MP-SPDZ 2 [p.7]
  • model: Logistic regression 24 [p.1], Linear regression 24 [p.1], SVM 5 [p.6], CNN 4 [p.29], k-means 3 [p.6], Decision tree 1 [p.6]
  • dataset: MNIST 7 [p.25], Credit card / German credit 1 [p.25]
  • hardware: Intel Xeon 1 [p.26]
  • technique: Three-party computation 58 [p.1], Semi-honest model 41 [p.2], Malicious model 37 [p.1], Two-party computation 31 [p.3], ReLU 6 [p.3], Convolution 6 [p.29], Matrix multiplication 5 [p.15]
  • task: Medical / healthcare 5 [p.1], Clustering 3 [p.6]
  • topic: Privacy-preserving machine learning 9 [p.1], Benchmark 4 [p.26], Homomorphic encryption 3 [p.6]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/schemes/s51-0_aby3-a-mixed-protocol-framework-for-machine-learning.pdf schemes 40 638.3 KB c4606cf65dca

Database key: aby3-a-mixed-protocol-framework-for-machine-learning · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set aby3-a-mixed-protocol-framework-for-machine-learning COLUMN VALUE