With the improvement of network infrastructures and advancement of IoT technologies, now it is desirable to perform computation at the edges, rather than sharing data with a central fusion center, which is privacy-intrusive. Both conventional (centralized) and distributed machine learning (ML) algorithms fail to address underlying challenges related to users’ privacy or capturing global knowledge of the whole network. To properly handle such challenges, a recently invented distributed ML technique, called Federated Learning was invented that shows us a pathway to construct a global model without exposing any user’s private data through on-device model training utilizing edge resources. However, FL may face various challenges due to the lack of a convenient mechanism to prepare a federated dataset, and also the heterogeneous nature of the resource-constrained agents such as systems, statistical and model heterogeneity. We developed a distributed sensing mechanism through which any federated agents can be triggered and activated for sensing the environment. That novel approach shows a pathway to carry out the FL process in a real-world environment. Following this, we developed an FL model, FedAR by monitoring agent activities and leveraging available local computing resources, particularly for resource-constrained IoT devices (e.g., mobile robots), to accelerate the learning process. Besides, we propose a tri-layer FL framework, FedPARL that helps resource-constrained FL agents consume less resources during training, and avoid untrustworthy and out-of-resource agents (e.g., low battery life) during agent selection for training and perform variable local epochs based on the agent’s resource availability. We perform model pruning to reduce the size of the agent model that is effective for an FL-IoT setting. Afterwards, we proposed a coupling of distillation and dynamic local task allocation technique through which we can effectively handle model and systems heterogeneity of FL-IoT environments. Further, we focused on applying our developed distributed ML algorithm to improve the resilience of critical infrastructures through knowledge exchange. We extended the work by proposing a novel technique, FedResilience, to handle weak critical infrastructure agents by enabling partial computational tasks from the resource-constrained agents and exchanging information without sharing any raw data. Besides, we implemented two other FL applications to recognize human activities and forecast customers’ financial distress in resource-constrained environments. From the experience we gathered over the past few years on FL, we conducted a comprehensive survey on FL, particularly for resource-constrained IoT environments that discussed the existing FL works, present challenges, and their potential solutions, applications, and future research directions in this domain.
