Existing Federated Learning (FL) frameworks do not support model’s evolvability in open Internet of Things (IoT) settings, where devices can belong to different users, and can contribute with data of different qualities. Moreover, the frameworks do not consider the participants’ trust issues in such dynamic settings. Towards addressing these challenges, this thesis will propose a distributed architectural approach that: 1) supports global models’ evolution in dynamic and open settings; 2) evaluates the trust scores of agents participating in FL rounds with respect to performances of their local models. To validate our approach of addressing the previously mentioned gaps in FL frameworks; we wrote the needed algorithms, developed a prototype, and ran experiments. The results we got prove the feasibility of our novel addition to the FL frameworks.