Performance Analysis of Machine Learning on Homomorphically Encrypted Data

The need for encrypted machine learning arises from the growing concern over data privacy and security in the era of big data and artificial intelligence. This study examines the performance of machine learning (ML) models on homomorphic encrypted data using the CKKS encryption scheme. We evaluated various models, including k-nearest neighbors (KNN), logistic regression (LR), and convolutional neural networks (CNN) on diverse datasets, including the Titanic, Pima Indians Diabetes, MNIST, and Fraud Detection datasets. Our experiments compare the performance of these models on encrypted and unencrypted data to assess the impact of encryption. The results demonstrate that while encryption introduces a slight performance degradation, the accuracy loss remains within acceptable limits. For instance, the logistic regression model's accuracy on the Titanic dataset was 82.72% for plain data, 81.36% for encrypted data with a polynomial degree of 4096, and 82.00% for a polynomial degree of 8192. The MNIST CNN model achieved 99.4% accuracy on plain data, 98.0% with 4096, and 99.0% with 8192. Our findings highlight the practical applicability of encrypted machine learning in preserving data privacy without significantly compromising performance, offering valuable insights for future research and development in secure machine learning systems.