Privacy-aware knowledge distillation for retinal scans de-identification through adversarial perturbations

Abstract

The rapid increase in retinal scan datasets over the past several years has played a pivotal role in advancing the deep learning approaches for retinopathy detection. However, given the retinal vessel pattern as an accurate biometric identifier, sharing this sensitive data raises critical ethical and legal considerations pertaining to individual re-identification and privacy violations. To mitigate these concerns, we propose a knowledge distillation-based approach that aims to disrupt identity recognition models by applying imperceptible adversarial perturbations to retinal scans while simultaneously preserving their utility for medical purposes. A multi-objective loss function is utilized for this approach, consisting of several terms that effectively guide the generator part of the student network to learn the effective perturbation needed to balance these conflicting goals. Extensive experiments and explainable analysis are conducted on three public fundus datasets using five deep architectures for identity recognition and retinopathy detection. With an average identity fooling rate of 94.67% and average retinopathy detection accuracy of 97.52% for multiple unseen models, the proposed approach outperforms state-of-the-art methods in balancing between medical utility and enhancing patient privacy.