The proliferation of Internet of Things (IoT) devices inte- grated with cloud computing infrastructures has created unprecedented opportunities for data-driven applications while simultaneously introduc- ing complex security vulnerabilities. This paper presents a unified deep learning framework that addresses the dual challenges of real-time threat detection and data integrity preservation in IoT-cloud ecosystems. Our proposed framework integrates a multi-layered neural architecture com- bining Convolutional Neural Networks (CNNs) for pattern recognition, Long Short-Term Memory (LSTM) networks for temporal analysis, and Graph Neural Networks (GNNs) for network topology understanding. The framework employs a novel adaptive threshold mechanism based on ensemble learning techniques to minimize false positives while main- taining high detection accuracy. Experimental validation on three di- verse IoT-cloud testbeds demonstrates superior performance with 97.8% threat detection accuracy, 2.3% false positive rate, and sub-second re- sponse times. The framework successfully identified advanced persistent threats, zero-day exploits, and data tampering attempts while preserv- ing system performance with minimal computational overhead. Our ap- proach represents a significant advancement in securing heterogeneous IoT-cloud environments through intelligent, adaptive, and scalable secu- rity mechanisms.