Professional paper

Application of homomorphic cryptography to machine learning models

Mária Krajčí ; Tehnički fakultet u Rijeci
Vesna Krajči ; Veleučilište u Rijeci

Abstract

Because of the growing need for processing medical data while ensuring its privacy, this paper explores the application of homomorphic encryption in testing machine learning models on medical images, using an approach based on the CryptoNets project. Homomorphic encryption enables the processing of encrypted data without decryption, thus allowing training and data analysis while preserving patient privacy. Testing the model on three separate medical data sets (ChestX-Det, Public Lung Dataset and JSRT Dataset) showed that the application of homomorphic encryption results in only a slight reduction in model accuracy. On the ChestX-Det dataset, which represents a multiclass classification problem, the F1 score remained at a satisfactory level, while the Public Lung Dataset and JSRT Dataset recorded slightly lower but acceptable accuracy levels. It was noted that the results on the JSRT dataset were poorer due to the smaller number of samples, which makes it difficult to learn more complex patterns. The paper concludes that homomorphic encryption provides a reasonable balance between privacy protection and model accuracy, opening opportunities for secure processing of sensitive medical data in the context of machine learning.

Keywords

homomorphic encryption, CryptoNets, machine learning

Hrčak ID:

328531

URI

https://hrcak.srce.hr/328531

Publication date:

19.12.2024.

Article data in other languages: croatian

Visits: 487 *