With the rapid advancement of generative models, deepfake creation has become increasingly sophisticated, posing serious security and privacy threats as it can be exploited for malicious purposes such as spreading misinformation, facilitating identity theft, and enabling other malicious activities. This highlights the urgent need for robust deepfake detection methods to identify manipulated content, particularly human faces. Existing deepfake detection methods primarily rely on encoder-decoder architectures, where image backbone models process visual information for classification while focusing on global features, resulting in inadequate detection accuracy and limiting their ability to capture fine-grained deepfake artifacts and localized forgeries. As generative models continue to advance, deepfake detection becomes increasingly challenging, particularly due to the emergence of fine-grained and localized manipulations that existing models struggle to detect due to their reliance on global features. We apply three network modules for multi-granularity feature extraction. The first module, the Global Feature Branch (GFB), extracts contextual global features to provide broad scene understanding, complementing fine-grained detection. The second module, the Generative Diffusion Feature Branch (GDFB), captures hierarchical multi-grained features, addressing the limitations of conventional backbone models and enhancing the detection of localized deepfake artifacts. The third module, the Sparse Auto-Encoder Branch (SAB), further refines fine-grained artifact detection by emphasizing subtle inconsistencies. Then, we aggregate these multi-grained features using a Feature Pyramid Network (FPN) to ensure a comprehensive representation. Finally, the fused representation is passed to a classifier, enabling robust differentiation between real and deepfake human face. Experimental results on three benchmark datasets prove the superior performance of MG-Det compared to existing state-of-the-art studies, where our model improves the best-performing study in the existing literature by 3.25% and 4.43% in Area Under the Curve (AUC) and accuracy (ACC), respectively.