Artificial Intelligence in Medicine and Privacy Preservation

The widespread applicability of medical artificial intelligence systems hinges on their development and validation on large, diverse, and representative datasets. So far, such datasets have only been able to be assembled through multi-institutional data sharing and aggregation. Such practices are however associated with legal, ethical, and technical challenges and scale poorly to multinational efforts. They furthermore potentially infringe on data ownership and complicate the enforcement of data governance measures. Privacy-preserving machine learning offers solutions to these challenges by implementing techniques for the decentralized training of algorithms on datasets without requiring direct access to the data or by offering guarantees of privacy protection during training and algorithm inference. This chapter presents the core techniques of secure and private artificial intelligence, which can serve to enable the training of algorithms on larger datasets and their provision to more people under provable assurances of privacy and ownership protection.