Federated machine learning in data-protection-compliant research

Paper by Alissa Brauneck et al : “In recent years, interest in machine learning (ML) as well as in multi-institutional collaborations has grown, especially in the medical field. However, strict application of data-protection laws reduces the size of training datasets, hurts the performance of ML systems and, in the worst case, can prevent the implementation of research insights in clinical practice. Federated learning can help overcome this bottleneck through decentralised training of ML models within the local data environment, while maintaining the predictive performance of ‘classical’ ML. Thus, federated learning provides immense benefits for cross-institutional collaboration by avoiding the sharing of sensitive personal data(Fig. 1; refs.). Because existing regulations (especially the General Data Protection Regulation 2016/679 of the European Union, or GDPR) set stringent requirements for medical data and rather vague rules for ML systems, researchers are faced with uncertainty. In this comment, we provide recommendations for researchers who intend to use federated learning, a privacy-preserving ML technique, in their research. We also point to areas where regulations are lacking, discussing some fundamental conceptual problems with ML regulation through the GDPR, related especially to notions of transparency, fairness and error-free data. We then provide an outlook on how implications from data-protection laws can be directly incorporated into federated learning tools…(More)”.