Federated Learning for Secure and Privacy-Preserving Medical Diagnostics

Federated learning is an emerging powerful approach for training ML models across decentralized data sources, such as in healthcare, without actually transferring the sensitive data to a central server. Although FL naturally reduces the risks of privacy since data are confined to local devices, it still can be highly vulnerable to sophisticated attacks such as model inversion, data poisoning, and inference attacks. Advanced privacy-preserving techniques that are nowadays studied for addressing the privacy and security concerns of this paper include differential privacy, homomorphic encryption, and secure aggregation. These, therefore, open the room for FL to be adopted as a solution in some of the privacy-sensitive domains. This paper conducts a thorough review of the techniques discussed and outlines a customized security framework for healthcare applications that allow patient data to be safeguarded without detrimental effects on model performance and accuracy. Conclusion: