DEVELOPING PRIVACY-PRESERVING FEDERATED LEARNING MODELS FOR COLLABORATIVE HEALTH DATA ANALYSIS ACROSS MULTIPLE INSTITUTIONS WITHOUT COMPROMISING DATA SECURITY

Federated learning is an emerging distributed machine learning technique that enables collaborative training of models among devices and servers without exchanging private data. However, several privacy and security risks associated with federated learning need to be addressed for safe adoption. This review provides a comprehensive analysis of the key threats in federated learning and the mitigation strategies used to overcome these threats. Some of the major threats identified include model inversion, membership inference, data attribute inference and model extraction attacks. Model inversion aims to predict the raw data values from the model parameters, which can breach participant privacy. The membership inference determines whether a data sample was used to train the model. Data attribute inference discovers private attributes such as age and gender from the model, whereas model extraction steals intellectual property by reconstructing the global model from participant updates. The review then discusses various mitigation strategies proposed for these threats. Controlled-use protections such as secure multiparty computation, homomorphic encryption and conidential computing enable privacy-preserving computations on encrypted data without decryption. Differential privacy adds noise to query responses to limit privacy breaches from aggregate statistics. Privacy-aware objectives modify the loss function to learn representations that protect privacy. Information obfuscation strategies hide inferences about training data.