Federated Learning enables distributed devices to collaboratively train a machine learning model while keeping data localized. However, scaling and securing client contributions remain significant challenges for achieving high-performing federated models. This project addresses these issues by developing a dynamic, on-demand crowdsourced FL framework that optimizes client participation through multi-criterion selection, secure aggregation using Bayesian inference, and incentivization mechanisms, thereby enhancing scalability, efficiency, and security in edge-based FL systems.