Adversarial Collaborative Learning on Non-IID Features

Federated learning has been a popular approach to enable collaborative learning on multiple parties without exchanging raw data. However, the model performance of federated learning may degrade a lot due to non-IID data. While most existing studies focus on non-IID labels, federated learning on non-IID features has largely been overlooked. Different from typical federated learning approaches, the paper proposes a new learning concept called ADCOL (Adversarial Collaborative Learning) for non-IID features. Instead of adopting the widely used model-averaging scheme, ADCOL conducts training in an adversarial way: the server aims to train a discriminator to distinguish the representations of the parties, while the parties aim to generate a common representation distribution. Our experiments on three real-world datasets show that ADCOL achieves better accuracy and is much more communication-efficient than state-of-the-art federated learning algorithms on non-IID features. More importantly, ADCOL points out a promising research direction for collaborative learning.