<p>Recent advances in generative modeling have enabled the creation of highly realistic deepfake facial images, posing significant risks to digital security, media integrity, and public trust. Although deep learning–based detection methods have achieved strong performance, they often suffer from limited cross-dataset generalization, sensitivity to manipulation-specific artifacts, and reduced interpretability. To address these limitations, this paper proposes a forensic-first hybrid deepfake face detection framework that integrates handcrafted local forensic descriptors with multi-CNN deep semantic representations. Specifically, manipulation-sensitive regions are captured using a Bag-of-Visual-Words (BoVW) model constructed from Histogram of Oriented Gradients (HOG) features extracted at salient keypoints detected via SURF, FAST, and BRISK. In parallel, high-level features are obtained from fine-tuned ResNet-50, MobileNet, and ShuffleNet models and fused at the feature level to capture complementary semantic information. The combined feature representation is classified using a Support Vector Machine (SVM), enabling stable decision boundaries and improved generalization. Extensive experiments on six benchmark datasets of varying scale and complexity demonstrate that the proposed approach consistently outperforms state-of-the-art methods, achieving up to 97.55% accuracy while maintaining robustness under cross-dataset and challenging forensic conditions. The results highlight the effectiveness of integrating explicit forensic features with deep representations to achieve a robust, interpretable, and generalizable solution for deepfake face detection.</p>