Federated learning (FL) is a promising technology wherein edge devices/clients collaboratively train a machine learning model under the orchestration of a central server. However, due to the inherent data heterogeneity among clients, local datasets on individual clients often exhibit class imbalance, i.e., samples from \textit{majority} classes vastly outnumber those from \textit{minority} classes. This imbalance significantly diminishes the performance of the trained model. To understand why, we first closely examine the output probability distribution clusters of the local deep neural networks (DNNs) in the probability space over the label set, and observe that for class imbalanced datasets, FL has two interesting phenomena: (1) dispersion problem---clusters corresponding to minority classes tend to disperse; and (2) gravity problem---clusters corresponding to minority classes are drawn toward those of majority classes. To overcome these two problems, we then introduce information quantities into FL, propose a new information theoretic loss function for FL, and develop a new FL framework called Fed-IT. It is shown that Fed-IT significantly outperforms previous counterparts, while maintaining client privacy.