Due to the social and cyber security threats posed by Deepfake technology, facial forgery detection has gradually become a focal point of public concern. However, existing facial forgery detectors often suffer from overfitting problems. We observe that this performance bottleneck mainly stems from the reliance of current detectors on a single backbone architecture: CNN-based architectures suffer from “regional dependency” caused by channel redundancy, leading to overfitting to method-specific forgery features in the training data. In addition, forgery traces are widely distributed across spatial scales, and a single architecture, limited by its receptive field, is unable to capture both detailed and structural forgery clues simultaneously, thereby leading to limited scale-aware capability and affecting generalization performance. To tackle these issues, we propose a Multi-dimensional Fusion Attention Network (MFA-Net), integrating a Multi-dimensional Feature Extraction Module (MFEM) and a Dual Attention-Driven Feature Fusion Mechanism (DAFFM) for comprehensive modeling of forgery features. MFEM combines four complementary backbones with varied architectures and supervision paradigms, mitigating regional dependency and enhancing scale representation. DAFFM fuses Content-Guided Attention (CGA) and Class-Specific Residual Attention (CSRA) to boost sensitivity to forged regions and stabilize multi-scale feature capture. Extensive experiments on public datasets demonstrate that MFA-Net achieves superior detection accuracy and significantly improves generalization in complex real-world scenarios.