Costin-Alexandru Deonise1, Ana-Maria Roangheși1, Liviu-Ionuț Gheorghe1, Emil Simion1, Bogdan-Costel Mocanu1, Dinu Țurcanu2, Florin Pop1,3,4
Abstract. The integration of Machine Learning into sectors that involve sensitive data is very often obstructed by privacy. This is dangerous, as the release of raw inputs or intermediate gradients leaks significant amounts of confidential information. We propose a controlled empirical comparison of BFV and CKKS for encrypted inference on linear and shallow convolutional models. Homomorphic Encryption tackles this challenge by performing computation directly in the encrypted domain, ensuring data remains protected throughout the entire inference pipeline. This work investigates the practical feasibility of HE-enabled inference for text classification and image recognition tasks. We implement and benchmark Logistic Regression and simplified CNNs using the BFV and CKKS encryption schemes. Our results outline the performance limitations of fully homomorphic inference, while demonstrating how optimizations – including quantization, batching, and parameter selection-can improve practical viability and extend applicability beyond purely research settings. This would therefore enhance the tuning that can be done to make such models viable outside the research domain.
Keywords: Homomorphic Encryption, BFV, CKKS, Logistic Regression, CNN.
DOI 10.56082/annalsarsciinfo.2025.2.50
1National University of Science and Technology POLITEHNICA Bucharest (e-mail: costin.deonise@upb.ro, ana_maria.roanghesi@stud.acs.upb.ro, liviu.gheorghe1802@stud.acs.upb.ro, bogdan_costel.mocanu@upb.ro, emil.simion@upb.ro, florin.pop@upb.ro)
2Technical University of Moldova (email: dinu.turcanu@utm.md)
3National Institute for Research & Development in Informatics – ICI Bucharest
4Academy of Romanian Scientists
PUBLISHED in Annals of the Academy of Romanian Scientists Series on Science and Technology of Information, Volume 18, No2
ISSN PRINT2066 – 2742 ISSN ONLINE 2066-8562 
Description | Editorial Board | Instructions for authors | Template 
| Archive