Privacy-Preserving Ridge Regression Over Encrypted Data Under Multiple Keys

With the increase of private data being collected by data owners, it has been a trend for data owners to store the data on cloud computing platforms. The huge amounts of data in cloud servers bring fresh development opportunities to machine learning, which is applied to build a high-quality machine learning model based on a large training dataset. However, to ensure the privacy of data and facilitate retrieval, data owners often upload encrypted data under their public keys. But it creates new challenges for machine learning to learn a predictive model over these encrypted data under different keys. Most existing privacy-preserving machine learning works make contributions over one dataset or multiple datasets encrypted with the same public key, which limits the utilization of these schemes in practice. In this paper, we consider a more practical scenario as the secure learning algorithm involves multiple data sources encrypted with different public keys. In this setting, we present a novel scheme for privacy-preserving ridge regression over arbitrarily-partitioned datasets to protect the confidentiality of raw data and the model. To achieve secure computing and the best performance, we construct privacy-preserving computation protocols using a linearly-homomorphic encryption (LHE) scheme and an additive homomorphic proxy re-encryption (PRE) scheme. The security analysis shows that our scheme satisfies the predefined security requirements. The theoretical analysis and experimental evaluations on varying datasets demonstrate its practicability.