本地分布和全局分布有可能是不一致的——本地的少数类有可能是全局的多数类。出于隐私考虑,又不能暴露本地分布。
To overcome the aforementioned challenges, we propose a constrained learning formulation to handle the class imbalance issue in FL while accounting for both heterogeneity and privacy. In brief, we impose constraints on the standard FL formulation so that the empirical loss on every client should not overly exceed the average empirical loss. Such constraints are shown to force the classifier to account for all classes equally and hence mitigate the detrimental effect of class imbalance, under a type of heterogeneous data configuration that captures the mismatch between the local and global imbalance.