Distributed SVM Training With Secret Sharing for Privacy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing distributed support vector machine (SVM) methods face challenges in balancing computational overhead and security, with high-security methods having high computational overhead and high-efficiency methods lacking security, while also failing to address privacy leakage during collaborative training across multiple machines and data sources.
Innovation Solution
A distributed SVM privacy-preserving method using secret sharing and secure multi-party computation, where entities exchange divided information iteratively and solve sub-problems locally, employing ADMM algorithm and methods like damped Newton's method to achieve convergence, ensuring privacy and accuracy without excessive computational overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If homomorphic encryption is used for privacy-preserving computation, then security is improved, but computational overhead increases excessively
Solution Approach 1:
The patent segments the secret parameters into multiple shares distributed among different parties. Instead of using homomorphic encryption on entire datasets, the ADMM algorithm operates on segmented secret shares locally, reducing the computational burden while maintaining privacy through secret sharing mechanisms.
Solution Approach 2:
The patent introduces secret sharing as an intermediary mechanism between the need for privacy protection and efficient computation. Rather than directly applying heavy cryptographic operations, the system uses secret shares as intermediaries that enable collaborative computation without exposing sensitive data, thus reducing overall computational overhead.
2Reliability
If perturbation method with differential privacy is used, then privacy is preserved, but convergence speed slows down and accuracy deteriorates
Solution Approach 1:
The patent performs preliminary secret sharing of parameters before the iterative optimization process begins. By pre-processing the data into secret shares and establishing the distributed structure in advance, the system avoids the need for perturbation methods during iteration, thereby maintaining both privacy protection and fast convergence characteristics of the ADMM algorithm.
3Productivity
If data is shared across multiple entities for distributed training, then processing capability is improved, but privacy leakage risk increases
Solution Approach 1:
The patent segments both the data and the secret parameters into shares distributed across multiple entities. Each party holds only a portion of the secret shares and can perform local computations without accessing complete sensitive information, thus enabling distributed processing while preventing privacy leakage through the properties of secret sharing.
Solution Approach 2:
The patent enables each distributed entity to perform local optimization using its own data and local secret shares. Each node processes information locally without needing to reveal its data or intermediate states to other parties, maintaining privacy while leveraging distributed computational resources through localized processing.
Data Source
AI summary
A distributed support vector machine privacy-preserving method includes: dividing a secret through secret sharing among all participating entities, iteratively exchanging a part of the information divided by the participating entities, and solving sub-problems locally; performing an iteration until a convergence is reached to find a global optimal solution; and in consideration of the generality of the privacy-preserving method, adopting a privacy-preserving method based on a vertical data distribution and a privacy-preserving method based on a horizontal data distribution, respectively; wherein the participating entities do not trust each other, and interact through a multi-party computation for local training. The method is applied to an honest-but-curious scenario, and uses the idea of data division to perform local computation through the interaction of part of the data among users to finally reconstruct the secret to preserve data privacy.