The rapid expansion of smart cities has led to the integration of Artificial Intelligence (AI) at the edge, enabling real-time decision-making for intelligent urban infrastructure. However, conventional centralized AI models pose critical challenges, including data privacy risks, security vulnerabilities, and high computational overhead. This paper investigates Federated Learning (FL) as a transformative paradigm to enhance security, privacy, and efficiency in edge AI systems for smart cities. Unlike traditional AI training methods, to cyber threats while ensuring compliance with data protection regulations. To address key challenges in heterogeneous smart city environments, we propose a hybrid optimization framework integrating differential privacy, secure multi-party computation (SMPC), and blockchain-based authentication. This approach strengthens resilience against adversarial attacks while ensuring secure model updates. Additionally, we introduce an adaptive aggregation mechanism, which dynamically adjusts model updates based on device reliability, data distribution, and network conditions, optimizing both learning efficiency and energy consumption in edge AI networks. Extensive experimentation on real-world smart city datasets demonstrates that the proposed framework enhances model accuracy, robustness, and privacy preservation compared to conventional AI approaches. Our findings establish Federated Learning as a cornerstone for secure, scalable, and privacy-aware AI in smart cities, facilitating trustworthy deployment of intelligent urban infrastructure. This research provides valuable insights for policymakers, researchers, and industry professionals, paving the way for next-generation AI-driven smart cities with enhanced security, privacy, and efficiency.