Linear Support Vector Regression in Cloud Computing on Data Encrypted using Paillier Cryptosystem

Abstract

The use of linear support vector regression on private data in cloud computing must consider data privacy. Homomorphic encryption is an approach to address the problem. However, most of the existing approaches still use inefficient fully homomorphic encryption, in which both the training data and the testing data must be encrypted using the same public key. This leads to the repetition of the training process. The problem is addressed in this paper by applying partially homomorphic encryption using Paillier cryptosystem. Operations in linear support vector regression are modified so that they can be applied to process encrypted data. The model is used to predict the motor and total UPDRS (Unified Parkinson's Disease Rating Scale) scores. To assess the performance of the model, the MRSE (Mean Root Square Error) of the prediction on encrypted data is then compared with the MRSE of the prediction on unencrypted data. The evaluation shows that the MRSE of the prediction on encrypted data is exactly the same as that on unencrypted data, which proves that the modification on the operations in linear support vector regression has been done correctly.