Detection of deep fake has also become a top priority in digital security and media integrity domains, particularly with the widespread presence of sophisticated generative adversarial network (GAN)-based methods. The current paper presents a novel hybrid detection system that combines graph-based formulations, a diffusion-osmosis filter specifically developed for attenuation of adversarial noise, and a graph convolutional network (GCN) classifier. The proposed system is extensively evaluated on diverse benchmark datasets: DFDC, Celeb A, Face Forensics++ (FF++), and CIFAR-10. By converting the images into a graph structure using an initial processing layer of a diffusion-osmosis filter, the system efficiently nullifies adversarial perturbations induced by GANs, thereby improving the robustness of the detection pipeline. The GCN classifier utilizes the graph-structured data to attain state-of-the-art performance in real-synthetic media discrimination. Our experimental findings indicate that the proposed framework achieves detection accuracy between 98.71% and 99.87% on the tested datasets. These evidences suggest that the proposed framework is capable of addressing a wide range of deep fake cases, including high-quality face manipulations and refined adversarial attacks. Produces a state-of-the-art scalable detection system which is immune to noise and may have potential for use in applications for content authentication, forensic analysis, and digital trust protection. The presented approach establishes a sound bench for future development of hybrid deep fake detection systems.