Quantitative Dissection of Spatial, Temporal, and Frequency Features in Deepfake Detection

Deepfake detection models have achieved remarkable accuracies yet remain opaque in their decision processes. In this study, we dissect the feature hierarchies exploited by four state-of-the-art deepfake detectors—Xception, ResNet50, EfficientNet-B4, and Vision Transformer (ViT), using advanced explainability methods (Grad-CAM, LIME, SHAP) on the FaceForensics++ dataset across various compression levels. Our experiments demonstrate that localized spatial artifacts, temporal discontinuities, and frequency-based anomalies are critical for discrimination. For instance, occluding the mouth region reduces Xception’s accuracy from 92.5% to 55.3% (a 37.2% drop), while blurring the eye region induces a 28% accuracy reduction for Xception and 35% for EfficientNet-B4. Temporal perturbation via inter-frame Gaussian smoothing decreases EfficientNet-B4’s performance from 95.2% to 64.7% (a 30.5% drop), underscoring its reliance on micro-motion cues. Moreover, Fourier analysis reveals that real videos exhibit 31.2% high-pass energy in the forehead region versus 22.8% in deepfakes, and filtering these features lowers EfficientNet-B4’s accuracy from 94.1% to 66.4%. These quantifiable metrics confirm that deepfake models integrate spatial, temporal, and frequency-domain features for classification. Our work provides a reproducible framework for interpretable deepfake detection, paving the way for more robust cross-dataset generalization and enhanced trustworthiness in media forensics.