Healthcare Privacy-Preserving Federated Transfer Learning using CKKS-Based Homomorphic Encryption and PYHFEL Tool

Abstract

 Digitization of healthcare data has shown an urgent necessity to deal with privacy concerns within
the field of deep learning for healthcare organizations. A promising approach is federated transfer learning,
enabling medical institutions to train deep learning models collaboratively through sharing model parameters rather
than raw data. The objective of this research is to improve the current privacy-preserving federated transfer
learning systems that use medical data by implementing homomorphic encryption utilizing PYthon for
Homomorphic Encryption Libraries (PYFHEL). The study leverages a federated transfer learning model to classify
cardiac arrhythmia. The procedure begins by converting raw Electrocardiogram (ECG) scans into 2-D ECG
images. Then, these images are split and fed into the local models for extracting features and complex patterns
through a finetuned ResNet50V2 pre-trained model. Optimization techniques, including real-time augmentation
and balancing, are also applied to maximize model performance. Deep learning models can be vulnerable to
privacy attacks that aim to access sensitive data. By encrypting only model parameters, the Cheon-Kim-Kim-Song
(CKKS) homomorphic scheme protects deep learning models from adversary attacks and prevents sensitive raw
data sharing. The aggregator uses a secure federated averaging method that averages encrypted parameters to
provide a global model protecting users’ privacy. The system achieved an accuracy rate of 84.49% when evaluated
using the MIT-BIH arrhythmia dataset. Furthermore, other comprehensive performance metrics were computed to
gain deeper insights, including a precision of 72.84%, recall of 51.88%, and an F1-score of 55.13%, reflecting a
better understanding of the adopted framework. Our findings indicate that employing the CKKS encryption scheme
in a federated environment with transfer cutting-edge technology achieves relatively high accuracy but at the cost
of other performance metrics, which is lower in the encrypted settings when compared to the plain one, an
acceptable trade-off to ensure data privacy through encryption with achieving an optimal model performance.