Secure machine learning on encrypted data using homomorphic encryption techniques

Privacy preservation has become a critical challenge in modern machine learning applications, especially in sensitive domains such as healthcare, finance, and cloud-based services, where confidential data must be protected. This paper aims to develop a secure machine learning framework using Homomorphic Encryption (HE), specifically the CKKS scheme, to enable computation on encrypted data without requiring decryption. The proposed methodology integrates Logistic Regression, Encrypted Fully Connected Neural Networks (EncFCNN), and Encrypted Convolutional Neural Networks (EncCNN) to perform both training and inference directly in the encrypted domain. The encrypted data and model parameters are processed securely on an untrusted server while maintaining complete data confidentiality. Experimental evaluation is conducted using the Pima Indian dataset and the MNIST dataset to compare encrypted and non-encrypted model performance. The findings show that the encrypted logistic regression model achieves an accuracy of 0.7375, which is comparable to plaintext training, while EncCNN achieves 100% accuracy on the test set with acceptable computational overhead. The results confirm that homomorphic encryption provides strong privacy preservation without significantly affecting model accuracy, making it highly suitable for secure real-world machine learning applications.