Background: Privacy-preserving machine learning (PPML) is increasingly important as organizations seek to leverage sensitive data while complying with data protection regulations. Homomorphic encryption schemes, such as CKKS and Paillier, enable machine learning on encrypted data, ensuring data confidentiality. However, these encryption methods introduce computational overhead that can affect the efficiency and scalability of machine-learning models. This thesis investigates the trade-offs between encryption strength, computational efficiency, and resource utilization in machine learning models trained on encrypted data.

Objectives: The primary objective of this research is to evaluate the impact of CKKS and Paillier encryption on key machine learning metrics—model accuracy, training and testing times, and resource utilization (CPU and RAM consumption)—across three machine learning algorithms: Linear Regression, Logistic Regression, and SVM. This study provides insights into how organizations can optimize their machine-learning workflows while ensuring data privacy.

Methods: We conducted a series of experiments using both CKKS and Paillierencryption schemes to train and test machine learning models on datasets of varying dimensionality (low, medium, and high attributes and instances). Results were analyzed in terms of accuracy, training/testing time, and resource consumption, providing a comparative analysis of the two encryption methods across different algorithms.

Results: CKKS and Paillier encryption preserved model accuracy across all tested algorithms. However, CKKS consistently outperformed Paillier in terms of training and testing times, making it more suitable for real-time applications. Paillier provided higher precision in Linear Regression but resulted in significantly longer training times, especially with high-dimensional datasets. CKKS also exhibited higher RAM consumption, whereas Paillier was more memory-efficient, albeit at the cost of longer processing times. 

Conclusions: This study provides a comprehensive evaluation of how homomorphic encryption impacts machine learning performance. CKKS is better suited for organizations requiring quick model updates and real-time predictions, while Paillier is preferable for applications requiring higher precision and lower memory usage. The findings offer actionable insights for balancing data privacy with computational efficiency in privacy-preserving machine learning implementations.