The growth of smart cities, combined with the widespread adoption of Internet of Things (IoT) devices, has resulted in an exponential surge in data output. With the increase of data, protecting privacy is critical, especially when dealing with sensitive information like air quality accuracy, traffic accuracy, and energy consumption accuracy. Federated learning (FL) offers a solution by allowing collaborative model training without centralized data, hence protecting privacy. However, typical FL approaches confront issues such as communication overhead, data heterogeneity, and security threats. This study presents a novel federated learning architecture that addresses these difficulties in smart cities. Our solution incorporates the artificial intelligence-enabled Tunicate Swarm Algorithm (TSA) into the FL process, improving model optimization and communication efficiency while preserving strong privacy protections. The TSA method is used to optimize the feature selection process for FL models. The framework is tested against an IoT dataset to forecast the accuracy of energy, traffic, and air quality. The results show that the suggested framework provides better and faster convergence than previous techniques, while successfully minimizing privacy and security threats.