A Comprehensive Survey of Federated Open-World Learning

The rapid development of large-scale, diverse data and machine learning (ML) technologies has facilitated the rise of intelligent applications across various sectors. However, concerns over privacy and security risks associated with data collection and ML model training have prompted the emergence of Federated Learning (FL), a distributed machine learning paradigm that ensures privacy-preserving capabilities through collaborative model training. Initially applied to Google's Gboard, FL has since found widespread adoption in domains such as intelligent transportation, recommendation systems, and healthcare. Despite its success, existing FL models primarily focus on optimizing performance while safeguarding privacy, often overlooking the collaborative group's ability to adapt to environmental changes. Drawing parallels to human societies, which effectively adapt to both individual and collective changes, we propose that FL's adaptation to dynamic environments can enhance its ability to support real-world applications. This paper introduces the concept of Federated Open-World Learning (FOWL), a framework that enables FL to not only respond to changes but learn patterns from them, addressing challenges such as participant variability, multi-task learning, and catastrophic forgetting. We provide an evolving view of FL techniques, discuss the shift towards open-world conditions, and compare existing methods for implementing FOWL. Finally, we highlight a number of key challenges and potential future research directions for advancing federated learning in dynamic, open-world environments.