Lattice-Based Cryptography for Privacy Preserving Machine Learning
Master thesis
Permanent lenke
https://hdl.handle.net/11250/3090547Utgivelsesdato
2023Metadata
Vis full innførselSamlinger
- Studentoppgaver (TN-IDE) [823]
Sammendrag
The digitization of healthcare data has presented a pressing need to address privacyconcerns within the realm of machine learning for healthcare institutions. One promisingsolution is Federated Learning (FL), which enables collaborative training of deep machinelearning models among medical institutions by sharing model parameters instead of rawdata. This study focuses on enhancing an existing privacy-preserving federated learningalgorithm for medical data through the utilization of homomorphic encryption, buildingupon prior research.
In contrast to the previous paper this work is based upon by Wibawa, using a singlekey for homomorphic encryption, our proposed solution is a practical implementationof a preprint by Ma Jing et. al. with a proposed encryption scheme (xMK-CKKS)for implementing multi-key homomorphic encryption. For this, our work first involvesmodifying a simple “ring learning with error” RLWE scheme. We then fork a popular FLframework for python where we integrate our own communication process with protocolbuffers before we locate and modify the library’s existing training loop in order to furtherenhance the security of model updates with the multi-key homomorphic encryptionscheme. Our experimental evaluations validate that despite these modifications, ourproposed framework maintains robust model performance, as demonstrated by consistentmetrics including validation accuracy, precision, f1-score, and recall.