Enhanced multi-key privacy-preserving distributed deep learning protocol with application to diabetic retinopathy diagnosis

Abstract

In this work, privacy-preserving distributed deep learning (PPDDL) is re-visited with a specific application to diagnosing long-term illness like diabetic retinopathy. In order to protect the privacy of participants datasets, a multi-key PPDDL solution is proposed which is robust against collusion attacks and is also post-quantum robust. Additionally, the PPDDL solution provides robust network security in terms of integrity of transmitted ciphertexts and keys, forward secrecy, and prevention of man-in-the-middle attacks and is extensively verified using Verifpal. Proposed solution is evaluated on retina image datasets to detect diabetic retinopathy, with deep learning accuracy results of 96.30%, 96.21% and 96.20% for DDL, DDL + SINGLE and DDL + MULTI scenarios respectively. Results from our simulation indicate that accuracy of the PPDDL is maintained while protecting the privacy of the datasets of participants. Our proposed solution is also efficient in terms of the communication and run-time costs.