Three-Party Computation for Privacy-Preserving Machine Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current privacy-preserving machine learning techniques are inefficient in multi-party computation settings, particularly in three-party models, due to limitations in data sharing and security against malicious attackers, leading to performance gaps compared to plaintext training.
Innovation Solution
Implementing a three-party computation model where private data is secret-shared among three training computers, allowing different parts of a data item to be stored on each computer, enabling efficient weight determination and computation through truncation and conversion techniques, such as delayed resharing and oblivious transfer, to maintain data privacy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If secret-sharing is used among three training computers to preserve data privacy, then security against malicious attackers is improved, but computational efficiency deteriorates
Solution Approach 1:
The patent divides private data into secret-shares distributed across three training computers, where each computer holds only a portion of the data. This segmentation enables security against malicious attackers while maintaining computational efficiency through localized processing of share data rather than centralized sensitive data handling.
Solution Approach 2:
The patent introduces share data as an intermediary representation that allows computations to be performed on encrypted portions of the original data. Training computers process share data through secret-sharing protocols, enabling efficient computation without directly accessing the underlying private data, thus resolving the contradiction between security and efficiency.
2Loss of information
If data is secret-shared among multiple training computers, then data privacy is improved, but training speed deteriorates
Solution Approach 1:
The patent performs preliminary secret-sharing of data before the training process begins. By pre-dividing data into shares and distributing them across three computers, the system establishes the secure computation framework in advance, allowing subsequent training operations to proceed efficiently without repeated privacy-preserving overhead during the actual training iterations.
Solution Approach 2:
The patent transforms the training process by changing the representation parameter of data from plaintext to secret-shares. This parameter change enables privacy-preserving computations while optimizing training speed through efficient share-based arithmetic operations and reduced communication overhead compared to traditional secure multi-party computation approaches.
3Reliability
If three-party computation is implemented with malicious security, then security guarantees are improved, but performance gap from plaintext training increases
Solution Approach 1:
The patent implements a dynamic three-party computation system where share data can be efficiently transformed and processed during training. The system dynamically adjusts the computation workflow to minimize communication rounds and computational overhead, achieving malicious security guarantees while reducing the performance gap from plaintext training through optimized interactive protocols.
Solution Approach 2:
The patent replaces traditional mechanical secure computation mechanisms with share-based cryptographic operations. By substituting complex secure multi-party computation protocols with efficient share arithmetic and transformation operations, the system achieves strong security guarantees while significantly improving performance and reducing the gap from plaintext training efficiency.
Data Source
AI summary
Methods and systems according to embodiments of the invention provide for a framework for privacy-preserving machine learning which can be used to obtain solutions for training linear regression, logistic regression and neural network models. Embodiments of the invention 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). Embodiments of the invention provide for efficient conversions between arithmetic, binary, and Yao 3PC, as well as techniques for fixed-point multiplication and truncation of shared decimal values. Embodiments also provide customized protocols for evaluating polynomial piecewise functions and a three-party oblivious transfer protocol.