Addressing Class Imbalance in Federated Learning via Collaborative GAN-Based Up-Sampling

Federated learning (FL) is an emerging learning framework that enables decentralized devices to collaboratively train a model without leaking their data to each other. One common problem in FL is class imbalance, in which either the distribution or quantity of the training data varies in different devices. In the presence of class imbalance, the performance of the final model can be negatively affected. A straightforward approach to address class imbalance is up-sampling, by which data of minority classes in each device are augmented independently. However, this up-sampling approach does not allow devices to help each other and therefore its effectiveness can be greatly compromised. In this paper, we propose FED-CGU, a collaborative GAN-based up-sampling strategy in FL. In FED-CGU, devices can help each other during up-sampling via collaboratively training a GAN model which augments data for each device. In addition, some advanced designs of FED-CGU are proposed, including dynamically determining the number of augmented data in each device and selecting complementary devices that can better help each other. We test FED-CGU with benchmark datasets including Fashion-MNIST and CIFAR-10. Experimental results demonstrate that FED-CGU outperforms the state-of-the-art algorithms.