A hybrid spatial-frequency attention-based algorithm using efficientnet for robust and interpretable deepfake...

The recent pace of generative media synthesis methods has greatly enhanced the credibility and accessibility of deepfake content, which causes serious risks to digital trust, authenticity of media, and forensic security. Existing deepfake detection methods are usually limited to either spatial domain visual cues or frequency domain artifacts, which leads to their limited robustness, poor generalization under realistic compression and poor interpretability. To overcome them, this paper will introduce a generalizable hybrid spatial-frequency deepfake detector, the proposed scheme combines both RGB-based visual representations with discrete cosine transform (DCT) frequency elements into a high-capacity convolutional network with attention-based refining. The suggested framework uses an EfficientNet-B7 backbone to identify rich hierarchical features and a convolutional block attention module (CBAM) to adaptively highlight information that is of interest to manipulation including spatial and channel-wise information. The early combination of spatial and frequency information allows the model to mutually exploit semantic inconsistencies and fine-scale high-frequency distortions added in the process of generating synthetic content. Comprehensive experiments of the FaceForensics + + C23 data set show that the proposed methodology has state-of-the-art performance with a ROC-AUC of 0.997, as well as high precision-recall balance and convergence of the training process. Further class separability is supported by feature-space analysis and prediction probability distributions and more complex CAM-based visualizations give significant forensic descriptions by identifying manipulation-prone regions of the faces. The high detection accuracy, the increased potential of generalization, and the greater interpretability are the factors that underline the efficiency of the suggested hybrid framework and confirm its appropriateness to the use in the field of the real-life deepfake forensics.