Deepfake detection remains a challenging task due to the diversity of forgery techniques and the distributional shifts between training and testing data. Existing methods, often framed as binary classification tasks, struggle with generalization, particularly in real-world scenarios involving unknown forgery types. To address this challenge, we propose DFMoE, a novel deepfake detection framework based on a Heterogeneous Mixture of Experts (HMoE) model. Our approach incorporates dynamic gating networks that adaptively select expert networks of varying capacities and scales according to the input characteristics, enabling precise identification of different types of forgeries. Leveraging a pre-trained face recognition model for multi-scale feature extraction, DFMoE combines expert specialization with adaptive data augmentation to enhance the detection of both known and unknown deepfake types. Experimental results show that our method significantly improves detection accuracy and robustness, offering a highly effective solution for deepfake detection across diverse scenarios.