面向云计算应用的用电负荷数据差分隐私保护方法

Differential privacy protection method of electrical load data towards cloud computing applications

随着云计算技术的发展,用户可以利用公共计算资源低成本、高效率地完成机器学习等大数据分析业务,但在提升计算效率和经济效益的同时,也面临隐私泄露风险。针对以机器学习即服务为代表的云计算中潜藏的用电负荷数据泄露问题,提出了一种差分隐私保护框架下基于时序生成对抗网络的用电负荷数据脱敏方法,通过使用满足差分隐私的脱敏合成数据替代原始敏感数据,从而有效阻止攻击者根据窃取的训练数据推断真实的敏感信息。引入瑞利差分隐私机制,在保留负荷数据统计学特征的前提下去除个体特征;在此基础上,采用循环神经网络作为生成对抗网络的生成器和判别器,捕获负荷数据的动态时间特性;同时,将自编码器与生成对抗网络相结合,进一步挖掘负荷数据的静态特征。通过理论推导证明了所提方法能够满足差分隐私要求,且可以对总隐私预算进行量化。数值实验结果表明,所提方法能保证隐私保护处理后用电负荷数据的隐私性和可用性。

With the development of cloud computing technology,users can use public computing resources to complete big data analysis services such as machine learning with low cost and high efficiency,which improves computing efficiency and economic benefits while facing privacy disclosure risk.Aiming at the users’ load data leakage problem hidden in the cloud computing represented by machine learning as a service,an electrical load data masking method based on time-series generative adversarial network under differential privacy protection is proposed.The synthetic data satisfying differential privacy is used to replace the original sensitive data so as to effectively prevent attackers from inferring real sensitive information from stolen training data.Rényi differential privacy mechanism is introduced to remove individual characteristics on the premise of keeping the statistical characteristics of load data.On this basis,the recurrent neural network is used as generator and discriminator of generative adversarial network to capture the dynamic time characteristics of load series.At the same time,the static characteristics of load series are mined by combining autoencoder with generative adversarial network.Theoretical derivation proves that the proposed method can meet the differential privacy requirements and the total privacy budget can be quantified.Numerical experiment results verify that the proposed method can ensure the privacy and availability of electrical load data after privacy protection processing.