Abstract

Homomorphic encryption enables computation directly on encrypted data, offering a promising solution for outsourced machine learning in privacy-critical settings, such as fraud detection in banking and medical diagnosis in healthcare. In particular, the Cheon–Kim–Kim–Song (CKKS) scheme is well-suited for machine learning inference, as it supports arithmetic over real-valued data. However, its practical deployment remains challenging due to the complexity of parameter selection, which requires balancing security, numerical precision, and computational efficiency. This thesis presents an automated, data-driven framework for CKKS parameter selection that incorporates both dataset characteristics and model structure to guide the generation and evaluation of candidate parameter configurations. Central to this approach is the use of test vectors representative of the underlying data, without requiring direct access to sensitive information. The parameter configurations are analyzed through a multi-objective trade- off framework, allowing users to reason about various performance metrics rather than relying on a single recommended solution. Experimental results demonstrate that the proposed framework identifies parameter configurations with optimal trade-offs while significantly reducing the computational cost of an exhaustive parameter search, and that representative test vectors effectively approximate real inference behavior. Overall, this work contributes to improving the usability of privacy-preserving machine learning by enabling automated, data-driven parameter selection while supporting exploration of application-specific trade-offs.

Advisor

Nord, Alex

Department

Computer Science

Disciplines

Artificial Intelligence and Robotics | Cybersecurity

Keywords

Homomorphic Encryption, Privacy-Preserving Machine Learning, Cryptography, Machine Learning, Cheon-Kim-Kim-Song (CKKS)

Publication Date

2026

Degree Granted

Bachelor of Arts

Document Type

Senior Independent Study Thesis

Download

Share

COinS
 

© Copyright 2026 Jonathan Rueffer