Cross-node knowledge transfer and generalisation based on federated meta-learning in fog computing

Fog computing environments comprise numerous heterogeneous and resource-constrained nodes, resulting in significant generalisation deficiencies in conventional machine learning models when addressing cross-node tasks characterised by non-independent and identically distributed data and limited labelling resources.This research offers a federated meta-learning architecture designed for fog computing, with the objective of facilitating rapid, privacy-preserving knowledge transfer and task adaptation between nodes.This framework combines the distributed model training paradigm of federated learning with the 'learning to learn' mechanism of meta-learning.We performed comprehensive tests on public datasets to simulate standard non-IID data settings in fog computing.Results indicate that, in comparison to conventional federated learning baselines, our methodology attains an average accuracy improvement of roughly 3.7% on novel tasks, while markedly enhancing model convergence and task adaptation efficiency.This suggests that the proposed approach significantly improves learning and generalisation capabilities for resource-limited nodes in fog computing settings while addressing unfamiliar tasks.