Privacy-Preserving Machine Learning for Health Institutes
Master thesis
Permanent lenke
https://hdl.handle.net/11250/3022598Utgivelsesdato
2022Metadata
Vis full innførselSamlinger
- Studentoppgaver (TN-IDE) [823]
Sammendrag
Medical data is, due to its nature, often susceptible to data privacy and security concerns.The identity of a person can be derived from medical data. Federated learning, onetype of machine learning technique, is popularly used to improve the privacy andsecurity of medical data. In federated learning, the training data is distributed acrossmultiple machines, and the learning process of deep learning (DL) models is performedcollaboratively. However, the privacy of DL models is not protected. Privacy attacks onthe DL models aim to obtain sensitive information. Therefore, the DL models should beprotected from adversarial attacks, especially those which utilize medical data. One of thesolutions to solve this problem is homomorphic encryption-based model protection. Thispaper proposes a privacy-preserving federated learning algorithm for medical data usinghomomorphic encryption. The proposed algorithm uses a Secure Multiparty Computation(SMPC) protocol to protect the deep learning model from adversaries. In this study, theproposed algorithm using a real-world medical dataset is evaluated in terms of the modelperformance.