Today’s advancement in Artificial Intelligence (AI) enables training Machine Learning (ML) models on the daily-produced data by connected edge devices. To make the most of the data stored on the device, conventional ML approaches require gathering all individual data sets and transferring them to a central location to train a common model. However, centralizing data incurs significant costs related to communication, network resource utilization, high volume of traffic, and privacy issues. To address the aforementioned challenges, Federated Learning (FL) is employed as a novel approach to train a shared model on decentralized edge devices while preserving privacy. Despite the significant potential of FL, it still requires considerable resources such as time, computational power, energy, and bandwidth availability. More importantly, the computational capabilities of the training devices may vary over time. Furthermore, the devices involved in the training process of FL may have distinct training datasets that differ in terms of their size and distribution. As a result of this, the convergence of the FL models may become unstable and slow. These differences can influence the FL process and ultimately lead to suboptimal model performance within a heterogeneous federated network.

In this thesis, we have tackled several of the aforementioned challenges. Initially, a FL algorithm is proposed that utilizes cluster analysis to address the problem of communication overhead. This issue poses a major bottleneck in FL, particularly for complex models, large-scale applications, and frequent updates. The next research conducted in this thesis involved extending the previous study to include wireless networks (WNs). In WSNs, achieving energy-efficient transmission is a significant challenge due to their limited resources. This has motivated us to continue with a comprehensive overview and classification of the latest advancements in context-aware edge-based AI models, with a specific emphasis on sensor networks. The review has also investigated the associated challenges and motivations for adopting AI techniques, along with an evaluation of current areas of research that need further investigation. To optimize the aggregation of the FL model and alleviate communication expenses, the initial study addressing communication overhead is extended to include a FL-based cluster optimization approach. Furthermore, to reduce the detrimental effect caused by data heterogeneity among edge devices on FL, a new study of group-personalized FL models has been conducted. Finally, taking inspiration from the previously mentioned FL models, techniques for assessing clients' contribution by monitoring and evaluating their behavior during training are proposed. In comparison with the most existing contribution evaluation solutions, the proposed techniques do not require significant computational resources.

The FL algorithms presented in this thesis are assessed on a range of real-world datasets. The extensive experiments demonstrated that the proposed FL techniques are effective and robust. These techniques improve communication efficiency, resource utilization, model convergence speed, and aggregation efficiency, and also reduce data heterogeneity when compared to other state-of-the-art methods.