Privacy-Preserving Technology Using Federated Learning and Blockchain in Protecting against Adversarial Attacks for Retinal Imaging.

Purpose: Collaboration provides valuable data for robust artificial intelligence (AI) model development. Federated learning (FL) is a privacy-enhancing technology that allows collaboration while respecting privacy via the development of models without raw data transfer. However state-of-the-art FL models still face challenges in non-independent and identically distributed (non-IID) health care settings and remain susceptible to privacy breaches. We propose an FL framework coupled with blockchain technology to address these challenges.

Design: Retrospective, multicohort study.

Main Outcome Measures: We evaluated our FL model performance in myopic macular degeneration (MMD) and OCT classification and compared our model against state-of the-art FL and centralized models.

Methods: A total of 27 145 images from Singapore, China, and Taiwan were used to design a novel FL aggregation method for the detection of MMD from fundus photographs and macular disease from OCT scans in feature distribution skew and label distribution imbalance scenarios. We further performed adversarial attacks (label flipping and clean label). As proof of concept, blockchain was incorporated into FL to demonstrate secure transfer of model updates across collaborating sites.

Results: Our FL model showed robust performance with an area under the curve (AUC) of 0.868 ± 0.009 for MMD detection and 0.970 ± 0.012 for OCT macular disease classification. In label flipping attack, our FL model had an AUC of 0.861 ± 0.019, similar to the centralized model (AUC 0.856 ± 0.015) and higher than other FL models (AUC 0.578-0.819). In clean label attack, our FL model had an AUC of 0.878 ± 0.006, which was comparable to the centralized model (AUC 0.878 ± 0.001) and superior to other state-of-the-art FL models with an AUC of 0.529 to 0.838. Simulation showed that the additional time with blockchain in 1 global epoch was approximately 5 seconds. The addition of blockchain to the FL framework was feasible with a minimal impact on model development time.

Conclusions: Our proposed FL algorithm overcomes the shortcoming of the traditional FL in non-IID situations and remains robust against adversarial attacks. The integration of blockchain adds further security during the transfer of model updates. Blockchain-enabled FL can be a trusted platform for collaborative health AI research.

Financial Disclosure(s): The author(s) have no proprietary or commercial interest in any materials discussed in this article.