Federated Learning (FL) has emerged as a paradigm shift enabling heterogeneous clients and devices to collaborate on training a shared global model while preserving the privacy of their local data. However, a common yet impractical assumption in existing FL approaches is that the deployment environment is static, which is rarely true in heterogeneous and highly-volatile environments like the Edge-Cloud Continuum, where FL is typically executed. While most of the current FL approaches process data in an online fashion, and are therefore adaptive by nature, they only support adaptation at the ML/DL level (e.g., through continual learning to tackle data and concept drift), putting aside the effects of system variance. Moreover, the study and validation of FL approaches strongly rely on simulations, which, although informative, tends to overlook the real-world complexities and dynamics of actual deployments, in particular with respect to changing network conditions, varying client resources, and security threats.

In this paper we make a first step to address these challenges. We investigate the shortcomings of traditional, static FL models and identify areas of adaptation to tackle real-life deployment challenges. We devise a set of design principles for FL systems that can smartly adjust their strategies for aggregation, communication, privacy, and security in response to changing system conditions. To illustrate the benefits envisioned by these strategies, we present the results of a set of initial experiments on a 25-node testbed. The experiments, which vary both the number of participating clients and the network conditions, show how existing FL systems are strongly affected by changes in their operational environment. Based on these insights, we propose a set of take-aways for the FL community, towards further research into FL systems that are not only accurate and scalable but also able to dynamically adapt to the real-world deployment unpredictability.