Personalized Online Federated Learning for IoT/CPS: Challenges and Future Directions

Abstract

In recent years, federated learning (FL) has emerged as a powerful paradigm for distributed learning thanks to its privacy-preserving capabilities. With the use of FL, a network of edge devices can make intelligent decisions without exposing their data to others. Despite its success, the traditional FL is not well suited to many practical applications such as those that involve the internet-of-things (IoT) or cyber-physical systems (CPS), where data access can be intermittent, and edge devices are semi-independent with device-specific dynamic behavior characteristics. Those devices are referred to here as semiindependent devices since they need to make decisions based on their own data and device characteristics, often independent of other devices and the information obtained from other devices in the network. Additionally, as new information becomes available, traditional FL must repeat the entire learning process and may not be able to provide timely and tailored solutions to participants. Personalized online FL, on the other hand, retains the collaborative and privacy-preserving aspects while learning in real time from intermittent data. It further enables devices to learn models customized to the device and the specific tasks it performs. In light of these reasons, personalized Online-FL is ideal for applications where the learning relies on heterogeneous data streams, and local optimization is beneficial. In this work, we want to bring attention to this new learning paradigm, present a few of the applications that could benefit from it, and highlight the principal challenges the research community faces in developing successful personalized Online-FL.