Several machine learning applications involve issues where privacy plays a special role. This includes cases in which secrecy applies to the training data (e.g., when it contains medical information) and those in which the algorithm itself is subject to protection because, for example, it reveals specific information about the training data. Our group is engaged in developing methods to protect such privacy. In particular, we are investigating solutions based on multiparty computation protocols technology. This technology allows calculations to be performed on distributed data so that the data is not revealed to outsiders. A related method we are also investigating in this context is the so-called homomorphic encryption, which further minimizes the need for interaction between the parties executing the algorithm. In addition, our research looks at possible applications of the so-called trusted execution environments to increase the privacy of the computations performed. A separate direction, which we are also investigating, is to ensure that the result of the computation does not reveal private information about the input data, using a technology called differential privacy. The method is based on appropriately adding random noise to the result, so that information about individual elements of the input data cannot be deduced from it.

The above-mentioned technologies have been developed in theoretical computer science for many years. The added value of our research comes from direct contact with IDEAS in machine learning researchers. These interactions allow us to tailor solutions to specific problems occurring in this field of research.