Inferring Class-Label Distribution in Federated Learning
Raksha Ramakrishna, György Dán · 2022
Federated Learning (FL) has become a popular distributed learning method for training classifiers by using data that are private to individual clients. The clients´ data are typically assumed to be confidential, but their heterogeneity and potential class-imbalance adversely impact the accuracy of the trained model. The class-imbalance may not be common knowledge or may even be confidential information itself. Thus, the inference of the class-label distribution of the training data is important both from a performance and from a privacy perspective. In this paper, we study the problem of class-label distribution inference from an adversarial perspective, based on model parameter updates sent to the parameter server. Firstly, we present conditions under which exact inference is possible. We then introduce four new methods to estimate class-label distribution in the general FL setting. We evaluate the proposed inference methods on four different datasets and our results show that they significantly outperform state of the art methods.