<p>Vehicular Ad Hoc Networks (VANETs) in urban environments are pivotal for intelligent transportation but remain highly susceptible to cyber-attacks, which exploit their decentralized and dynamic nature. Existing intrusion detection methods often suffer from excessive computational requirements, an inability to handle novel attacks, and challenges in maintaining real-time performance in dense traffic scenarios. These limitations restrict their practicality in urban VANET architectures. To address these challenges, a novel framework, Dynamic Gravitational tuned Multilayer Perceptron (DG-MLP), is introduced. The approach combines rule-based detection with a machine learning method optimized using Dynamic Gravitational Search Optimization (DGSO). DGSO facilitates efficient feature selection and hyperparameter tuning, ensuring the method remains computationally lightweight while achieving high detection accuracy. The system is evaluated on a diverse and publicly available VANET dataset encompassing multiple attack types and urban mobility scenarios. Preprocessing steps include data cleaning, Min-Max scaling, and feature extraction to standardize and enhance input quality. The hybrid architecture employs rule-based modules for immediate detection of known attack patterns, while the optimized Multilayer Perceptron (MLP) identifies novel and anomalous behaviors in real-time. The DG-LMP framework demonstrates a detection accuracy exceeding 97%, a false positive rate under 2%, and latency suitable for real-time deployment. These results emphasize its adaptability, scalability, and suitability as an efficient security solution for urban VANETs, providing robust protection against evolving cyber threats.</p>