Generative data augmentation with differential privacy for non-IID problem in decentralized clinical machine learning

Swarm learning (SL) is an emerging promising decentralized machine learning paradigm and has achieved high performance in clinical applications. By combining edge computing and a blockchain-based peer-to-peer network, SL overcomes the challenges of a centralized structure in federated learning. Although SL has achieved impressive results assuming independent and identically distributed (IID) data across participants, it experiences performance degradation when faced with non-IID data. To this end, we propose a novel framework called DP-SL-GAN, which leverages differential generative augmentation in SL. Which augments the non-IID data by generating synthetic data on each SL participant. DP-SL-GAN co-trains a differential generative model with the local target task model and integrates them periodically through a randomly elected coordinator in the SL network. Under the standard assumptions, we theoretically prove the convergence of DP-SL-GAN using stochastic approximations and provide the privacy guarantee for our method. Experimental results demonstrate that DP-SL-GAN outperforms state-of-the-art methods on three real world clinical datasets including Tuberculosis, Leukemia, COVID-19.