基于揭序加密的联邦决策树安全比较协议

A Secure Comparison Protocol for Federated Decision Trees Based on Order-Revealing Encryption

得益于联邦学习的发展,多个参与方得以越来越方便地在保护隐私的前提下协同训练多种机器学习模型.随着联邦学习训练方案的逐渐成熟,在这些模型上进行高效联邦预测任务的需求也日益受到关注.目前对联邦学习得到模型的应用的研究大都处于复用训练范式的范畴,而实际上这一做法造成了额外的开销,联邦预测任务存在着极大的效率提升空间.本文对基于树模型的外包预测服务进行研究.首先明确阐述了“联邦预测”任务,并指出设计高效解决方案面临着数据异构分布和用户数据动态变化的挑战.针对这些挑战,本文提出了新颖的安全比较协议OREC.OREC通过一个不可信的第三方的协助,使得两个参与方得以秘密地比较他们的私密值.所设计的协议基于揭序加密,其通过公开函数可以揭示出密文之间的顺序关系.为了增强其隐私性,我们进一步设计了一次一密机制并引入了混淆密码序列技巧.我们对协议进行了全面的安全性分析,并证明了其面对恶意第三方的鲁棒性.此外,我们还通过大量实验验证了所提出协议的高效性和可扩展性.

With the advancement of federated learning,multiple participants can collaboratively train diverse machine learning models while preserving privacy.As federated modeling techniques have been well-established,there is an increasing demand for efficient federated prediction tasks on these models.However,most existing research focuses on reusing training paradigms in fed-erated learning,which results in additional overhead.In this paper,we specifically investigate outsourced prediction services for tree-based models,such as gradient boosting trees and random forests.To address this challenge,we propose the OREC secure comparison protocol.This pro-tocol enables confidential values to be secretly compared with the assistance of an untrusted third party.By utilizing order-revealing encryption,the protocol unveils the order relationship between ciphertexts through public functions.To further enhance privacy,we introduce a one-time pad mechanism and obfuscated password sequences.Furthermore,we conduct a comprehensive secu-rity analysis to demonstrate the protocol's robustness against malicious third parties.Through extensive experiments,we validate its efficiency and scalability,highlighting its potential in prac-tical applications.