Privacy-preserving collaborative machine learning in biomedical applications

Machine learning (ML) algorithms are now widely used to tackle computational problems in diverse domains. In biomedicine, the rapidly growing amounts of experimental data increasingly necessitate the use of ML to discern complex data patterns. However, biomedical data is often considered sensitive, and the privacy of individuals behind the data is increasingly put at risk as a result. Traditional methods such as anonymization and pseudonymization are not always applicable and have limited effectiveness with respect to risk mitigation. Privacy researchers are actively developing alternative approaches to privacy protection, including strategies based on cryptography, such as homomorphic encryption and secure multiparty computation. This paper discusses recent advances in biomedical applications of these privacy techniques. We first review the key privacy techniques, then provide an overview of their applications in biomedical machine learning. Finally, we highlight the remaining challenges of current approaches and suggest directions for future work.