云计算环境下朴素贝叶斯安全分类外包方案研究

NAIVE BAYESIAN SECURITY CLASSIFICATION OUTSOURCING SCHEME IN CLOUD COMPUTING ENVIRONMENT

当前基于大数据环境的机器学习模型训练和使用模式正饱受争议,尤其在用户针对已训练模型输入特征实例得到分类结果的模型使用阶段。一方面用户不愿意在使用过程中暴露自己的输入数据及最终结果,另一方面模型拥有者迫切需要将分类业务外包给云服务器,同时不暴露模型的明文参数。基于此应用场景,提出一种基于同态加密技术及盲化技术的朴素贝叶斯安全分类外包方法,并在云计算环境下实现仿真。整个系统允许模型拥有者加密上传模型,用户与云服务器利用同态性质完成安全多方计算。在多个朴素贝叶斯分类实例上进行仿真,结果表明该方案在不降低分类准确率的前提下实现了针对训练模型、输入数据及分类结果的隐私保护。

The current training and usage patterns of machine learning models based on big data environments are controversial,especially in the model using phase of the classification results obtained by inputting feature instances for trained models.On the one hand,users are reluctant to expose their input data and final results.On the other hand,model owners urgently need to outsource the classification business to the cloud server without exposing the plaintext parameters of the model.Based on this application scenario,this paper proposes a Naive Bayesian security classification outsourcing system based on homomorphic encryption technology and model blinding technology,and implements simulation in cloud computing environment.The entire system allows the model owner to upload the encrypted model,and the user together with the cloud server use the homomorphic nature to perform secure multiparty computing.We simulate on a number of Naive Bayesian classification instances.The results show that the scheme achieves privacy protection for training models,input data and classification results without reducing the classification accuracy.