Towards Efficient and Privacy Preserving ECG Classification: Federated Transfer Learning Enhanced by CKKS-Based Homomorphic Encryption

In healthcare, maintaining both the accuracy and privacy of medical diagnoses collaboratively is a significant challenge. To the best of our knowledge, this research proposes the first end-to-end Privacy-Preserving Federated Transfer Learning (PPFTL) framework for 2-D ECG arrhythmia classification. Incorporating Transfer Learning (TL) narrows the gap between the encrypted and non-encrypted framework versions in the training step. It involves the transformation of raw Electrocardiogram (ECG) signals into 2-D ECG grayscale images. The dataset is disseminated after transformation to images and then fed as input into the local models, where MobileNetV2 serves as a feature extractor. The training process for each client incorporates data balance and augmentation techniques to improve the model's performance. Deep Learning (DL) models are subject to various privacy attacks to gain sensitive data. As a result, the Homomorphic Encryption Cheon-Kim-Kim-Song (HE-CKKS) scheme encrypts only model parameters to protect deep models from adversary attacks, preventing the sharing of sensitive raw data. Experimental results on the MIT-BIH Arrhythmia dataset achieved 88.12% accuracy. Incorporating HE-CKKS increased computation times by 1.08%, 1.27%, and 1.43% for 2, 3, and 4 clients, respectively.