Federated learning (FL) is attributed to training a machine learning (ML) model over a number of distributed devices while keeping all their training data localized. Under these settings, the edge devices perform computations on their local data before sending the required updates to the central server to improve the global model. This approach has shown great potential since hundreds of devices can potentially contribute to learning a single task without sharing their local data. Despite its success in many domains, current FL systems face significant challenges in scaling client participation and handling data heterogeneity, which impede the training of high-performing models. To address these limitations, this thesis argues that there’s a need for a dynamic learning platform where edge devices could volunteer to collaboratively learn a task through FL. It further proposes FedOnDemand, an on-demand crowdsourced FL framework that dynamically incorporates edge devices based on demand and availability. The research aims to optimize client participation and resource allocation in FL systems to ensure an efficient and scalable learning process. We present a novel client selection mechanism designed to optimize the contribution of clients based on their computational resources and data quality. By implementing a multi-criterion client selection protocol, the system dynamically selects clients based on their suitability for a given FL task. To secure this process, we incorporate attributebased access control measures, ensuring that client selection is both effective and secure. This approach not only enhances the quality of the contributions but also safeguards the integrity of the FL process. To manage data heterogeneity and improve model robustness, we model FL as a Bayesian process. Clients employ Stochastic Variational Inference (SVI) to approximate local posterior distributions, while the server utilizes Bayesian learning techniques to aggregate these updates, effectively managing uncertainty. Furthermore, we explore fairnessaware incentive mechanisms based on data valuation, ensuring clients are rewarded proportionally to their contributions. These mechanisms are designed to foster active and robust participation across diverse network environments. Empirical evaluations using benchmark datasets demonstrate significant improvements in convergence speed, model accuracy, and system scalability compared to traditional FL approaches. This research contributes to the field by providing a framework that enhances the operational efficiency of FL models and ensures greater participant engagement and system integrity. The implications of this study are far-reaching, potentially influencing future designs of ML systems that require decentralized data inputs across highly dynamic and privacy-sensitive environments.