Start of funding 01.01.2021

PRIMAL: Privacy-Preserving Federated Machine Learning

Prof. Dr. Alexandra Dmitrienko
University of Würzburg
Institute of Computer Science II

Prof. Dr. Farinaz Koushanfar
University of California, San Diego
Jacobs School of Engineering



The goal of this project is to improve security and privacy of Artificial Intelligence and Machine Learning methods. Privacy enhancements will be achieved through application of the Federated Learning (FL) – an approach where the training of models takes place locally, while the locally-trained models are then aggregated by the central instance before being re-distributed back to clients. FL helps to protect privacy of training data, since central collection of data is not required. However, FL methods by themselves are subjected to security and privacy risks. For instance, there are methods known to infer information about the training data from local model updates. Another known attack vector is poisoning attacks, where an attacker manipulates training data. In this project, we aim to solve open problems of FL methods and will design an improved method that will be both, privacy preserving and robust against data poisoning.