Secure ML Evaluation on Encrypted Data with Fully Homomorphic Encryption and Concrete ML

In healthcare sector, preserving patient data privacy is vital. The increasing popularity of machine learning (ML) models for classification tasks on sensitive medical datasets calls for the widespread use of strong encryption approaches to protect data. This work looks at employing homomorphic encryption to safeguard healthcare datasets for machine learning (ML) classification tasks. The performance of fully homomorphic encryption (FHE) with Cheon-Kim-Kim-Song(CKKS) on medical datasets is evaluated in terms of the time taken to encrypt/decrypt the datasets used for ML tasks and the time taken to perform fundamental operations like addition and dot product. In this study, ML models such as Logistic Regression, Support Vector Machine, Decision Tree, and XGBoost are trained on two datasets: heart disease prediction and breast cancer prediction, and then tested on unencrypted, quantized, and fully homomorphic encrypted data using scikit-learn and Concrete ML. The models are evaluated using accuracy, precision, recall and f1-score. This paper evaluates the practicality of using homomorphic encryption to maintain the confidentiality of data without significantly affecting performance of ML models, emphasising its potential benefits in healthcare.