Privacy of 5G Enabled Networks: Homomorphic Encryption based Privacy-Preserving Machine Learning
Master thesis
Permanent lenke
https://hdl.handle.net/11250/3090184Utgivelsesdato
2023Metadata
Vis full innførselSamlinger
- Studentoppgaver (TN-IDE) [866]
Sammendrag
Homomorphic encryption (HE) is a technique that allows computations to beperformed on encrypted data, just as if the data were unencrypted. This hasnumerous potential applications, such as sensitive medical data, mainly whenprivacy and anonymity are critical. HE can also be used in cases where multipleparties need to perform computations on shared data without revealing thedata to one another. One fascinating application of HE is in machine learning,specifically in a process known as federated learning (FL). FL is a cutting-edgemethod that is particularly useful in situations where privacy is essential, as iteliminates the need for data to be shared with a central server, as is the casewith traditional distributed machine learning models. However, privacy risksare associated with sharing model parameters, as inference attacks can obtainsensitive information. This issue can be addressed by encrypting the modelparameters with HE on the client side and aggregating the encrypted data.In this paper, we explore federated learning with homomorphic encryption toimprove the privacy of 5G networks. The results of our experiments show thatencryption has a minimal effect on the the predictive performance of the model,but leads to an increase in computation time by 587 %, 624 % and 679 % for2, 5, and 7 clients, respectively.