Federated Learning with Emerging New Class: A Solution Using Isolation-Based Specification

Abstract

This paper investigates an important classification problem in federated learning (FL), i.e., federated learning with emerging new class or FLENC, where instances with new class that has not been used to train a classifier, may emerge in the testing process. As based on the assumption of the fixed class set, the common FL classification approaches will face the challenge that the predictive accuracy severely degrades if they are used to classify instances with a previously unseen class in FLENC problem. To address this issue, we explore a new FL framework under the one-shot setting, called the federated model with specification. This framework is to learn the local model with the specification which can explain the purpose or specialty of the model, and learn the global federated model in only a single round of communication. Further, we introduce an implementation FLIK by employing a data-dependent isolation-based kernel as the specification. FLIK is the first attempt to propose a unified method to address the following two important aspects of FL: (i) new class detection and (ii) known class classification. We report evaluations demonstrating the effectiveness of our proposed method in FLENC problem.