Evaluation of Privacy-Preserving Support Vector Machine (SVM) Learning Using Homomorphic Encryption

The requirement for privacy-aware machine learning increases as we continue to use PII (personally identifiable information) within machine training. To overcome the existing privacy issues, we can apply fully homomorphic encryption (FHE) to encrypt data before they are fed into a machine learning model. This involves generating a homomorphic encryption key pair, where the public key encrypts the input data and the private key decrypts the output. However, there is often a performance hit when we use homomorphic encryption, so this paper evaluates the performance overhead of using an SVM (support vector machine) machine learning technique with the OpenFHE homomorphic encryption library. This uses Python and the scikit-learn library to create an SVM model, which can then be used with homomorphically encrypted data inputs and then produce a homomorphically encrypted result. The experiments include a range of variables, such as multiplication depth, scale size, first modulus size, security level, batch size, and ring dimension, along with two different SVM models, SVM-poly and SVM-linear. Overall, the results show that the two main parameters that affect performance are ring dimension and modulus size, and SVM-poly and SVM-linear show similar performance levels.