Fast homomorphic SVM inference on encrypted data

Kernel methods are popular machine learning methods that provide automated pattern analysis of raw datasets. Of particular interest is Support Vector Machines that are used to solve supervised machine learning problems in many areas such as business, finance and healthcare. Nowadays, complex computations and data analytic tasks can be outsourced to specialized third parties. However, data owners might be reluctant to share their data especially when it includes sensitive information. Therefore, a need for privacy-preserving machine learning applications cannot be overstated. We present FHSVM: a F ast H omomorphic evaluation of non-linear SVM prediction on encrypted data using Fully Homomorphic Encryption. We provide design, implementation and several algorithmic and architectural optimizations such as novel packing strategies and parallel implementation to achieve real-time private prediction. We employed the CKKS FHE scheme to implement FHSVM under 128-bit security level. We evaluated FHSVM on a contemporary real-world large dataset compiled for anti-money laundering tasks in Bitcoin transactions. Empirical analysis demonstrates that homomorphic SVM prediction can be performed in 1.25 s on multi-core CPU platforms. In addition, FHSVM shows zero accuracy loss when compared to the non-privacy-preserving implementation. This shows that FHSVM is both computationally secure and fully utilizes the data.