On the impact of non-IID data on the performance and fairness of differentially private federated learning

Federated Learning enables distributed data holders to train a shared machine learning model on their collective data. It provides some measure of privacy by not requiring the data be pooled and centralized but still has been shown to be vulnerable to adversarial attacks. Differential Privacy provides rigorous guarantees and sufficient protection against adversarial attacks and has been widely employed in recent years to perform privacy preserving machine learning. One common trait in many of recent methods on federated learning and federated differentially private learning is the assumption of IID data, which in real world scenarios most certainly does not hold true. In this work, we empirically investigate the effect of non-IID data on node level on federated, differentially private, deep learning. We show the non-IID data to have a negative impact on both performance and fairness of the trained model and discuss the trade off between privacy, utility and fairness. Our results highlight the limits of common federated learning algorithms in a differentially private setting to provide robust, reliable results across underrepresented groups.