The integration of sixth-generation (6G) networks with Wireless Sensor Networks (WSNs) creates unprecedented opportunities for developing secure and scalable smart city infrastructures. However, the proliferation of heterogeneous devices and exponential data growth demand more robust security solutions. While existing hybrid deep learning approaches combining convolutional, recurrent, and attention-based architectures show promise in attack detection, they face limitations including high false-positive rates, inadequate modeling of topological dependencies, and vulnerability to adversarial attacks. This paper presents an enhanced intrusion detection framework that integrates Graph Neural Networks (GNNs) for structural dependency learning, cross-attention mechanisms for feature fusion, and stacked ensemble classification for improved decision reliability. Evaluated on Kitsune, 5G-NIDD, and CICIDS-2018 datasets, the framework demonstrates strong adaptability across heterogeneous traffic scenarios and complex attack vectors. Experimental results show remarkable performance with 99.95\% detection accuracy, consistent F1-scores above 99\%, significantly reduced false alarms, and enhanced adversarial resilience. These findings validate the framework's scalability and practical readiness for securing next-generation 6G-enabled smart city infrastructures.