Dr. Georgios Kaissis Reliable AI

AI in Medicine requires large, diverse, and representative datasets to train fair, generalisable and reliable models. However, such datasets contain sensitive personal information. Privacy-preserving machine learning bridges the gap between data utilisation and data protection by allowing the training of machine learning models on sensitive data while offering formal privacy guarantees. 

Our group focuses on applications of Differential Privacy to machine learning and deep learning, both on unstructured datasets such as images and on structured data such as tabular and graph databases. Moreover, we develop techniques for mitigating privacy-utility and privacy-performance trade-offs. Furthermore, we investigate attacks against collaborative machine learning protocols (such as federated learning) and develop defences against them.

Building trust in AI requires techniques to quantify the uncertainty of model outputs, incorporate domain expertise and optimally tackle training on small datasets. A further focus of our group is the development of probabilistic machine learning models which counteract the tendency of conventional models to produce overconfident and poorly calibrated predictions. We employ computational Bayesian techniques to train both statistical and deep learning algorithms and work on the intersection between probabilistic machine learning and privacy preservation.