Researches On Privacy-preserving Machine Learning Based On Multi-key FHE

In recent years,machine learning algorithms have been widely used in data analysis domain.Especially in the medical field,machine learning can assist doctors in diagnosis and treatment.It is not only improve efficiency,but also make full use of medical big data.For instance,clustering algorithms and classification algorithms are very powerful tools for breast cancer diagnosis.Meanwhile,with the incredibly increasing volume of medical data,machine learning algorithms are becoming more and more popular,and cloud computing will effectively help the storage and calculation of big data.However,cloud severs are generally considered not completely trust-worthy.Therefore,outsourcing medical data to cloud servers for storage and calculation will bring privacy risks.Once privacy medical information leaked,it may cause huge damage to data owners.So how to implement privacy-preserving machine learning under cloud environment is a current research hot-spot.Most existing studies only support single-data-owner setting,but data comes from distant devices or data owners in real application.Therefore,we adopt multi-key fully homomorphic encryption(Multi-Key FHE)to solve the above problem,our main research contents are:(1)We propose two schemes for privacy-preserving k-means clustering,PPK and PPOK.Both schemes support multi-data-owner setting,and are secure under the semi-honest model.(2)We propose two schemes for privacy-preserving SVM classifier,PPSC and PPOSC.Both schemes support several different kernel functions and multi-data-owner setting.Besides,PPSC and PPOSC are secure under the semi-honest model.(3)Through theoretical analysis and experimental simulation,we prove that both the PPK and PPSC are low cost in data owners' side,while the PPOK and PPOSC achieve completely outsourcing computation.