Abstract:
As privacy concerns grow alongside the increasing use of deep learning (DL)
in healthcare, Fully Homomorphic Encryption (FHE) presents a viable solution
for maintaining data confidentiality while enabling encrypted computation. This
study explores the implementation of encrypted Convolutional Neural Networks
(CNNs) for brain tumor detection within a secure client-server system. Two leading
FHE libraries—TenSEAL and Concrete ML—were evaluated in terms of classification
accuracy, runtime efficiency, and integration feasibility. The TenSEALbased
CNN preserved its plaintext accuracy (75%) after encryption, while the
Concrete ML model experienced a slight accuracy drop (from 82.5% to 75%). Despite
comparable runtime performance, TenSEAL’s consistent results and more
transparent parameter tuning made it the preferred choice for deployment. This
work contributes a novel use case of FHE in medical imaging beyond the standard
MNIST dataset and provides actionable insights for future implementations
of privacy-preserving machine learning in healthcare.
