Knowledge-Injected Federated Learning

With the development of artificial intelligence, people recognize that many powerful machine learning models are driven by large decentralized datasets of various data types. However, in many industryscale applications, training data is obtained and maintained by different data owners instead of centralized at the data center, and sharing data is often forbidden due to privacy requirements. Federated learning (FL) is an emerging machine learning framework in which multiple data owners (also referred to as clients) participate in collaboratively training a model without sharing their local data with each other [18, 33]. Another challenge with artificial intelligence is integrating domain knowledge into purely datadriven models, i.e., parameters of the model are learned through training data without any human engineering [8, 11]. For example, human know-how and craftsmanship, which may not be learnable from the training data, can be formulated as prediction models, and combing them with a purely data-driven model may boost its performance and reduce the risk of overfitting [10].