Federated Learning (FL) is a revolutionary privacy-preserving distributed learning framework that allows a small group of users to cooperatively build a machine-learning model using their own data locally. Smart cities are areas that can generate high volume and critical data, which has the potential to revolutionize federated learning. Nevertheless, it is highly challenging to select a trustworthy group of clients to collaborate in model training. The utilization of a random selection technique would pose many threats due to malicious clients’ targeted and untargeted attacks. Such vulnerability may cause attacks and poisoning in the produced model. To address this problem, we present a mutual trust client-server selection approach based on matching game theory and bootstrapping mechanisms for federated learning in smart cities. Our solution entails the creation of: (1) preference functions for federated servers and smart devices (i.e., IoT/IoV) that enables them to sort each other based on trust score, (2) light feedback-base technique that leverages the cooperation of multiple client devices to assign trust value to the newly connected federated server, and (3) intelligent matching algorithms consider trust preferences of both parties in their design. According to our simulation results, our technique outperforms the baseline selection approach VanillaFL in terms of increasing the trust level and hence the global accuracy of the federated learning model and optimizing the number of untrusted selected clients.