<![CDATA[Deepfake-enabled fraud caused nearly US$900 million in global losses during 2025, yet existing detection methods struggle with high-quality synthetic videos. Traditional CNN-based approaches fail to capture subtle global manipulations, while transformer-based models focus predominantly on spatial features, missing frequency-domain artifacts where manipulation traces are most evident. This research developed a dual-domain deepfake detection system combining spatial RGB analysis with Discrete Cosine Transform (DCT) frequency features, integrated into Data-Efficient Image Transformer (DeiT-Small) architecture. The proposed DeiT-DCT model processes four-channel inputs (RGB + DCT) enabling simultaneous learning of spatial textures and frequency anomalies. Training employed the AdamW optimizer with Cosine Annealing Warm Restart, Mixup augmentation with SoftTargetCrossEntropy loss, and domain-balanced sampling via WeightedRandomSampler. Evaluated on a combined dataset of five benchmarks (Celeb-DF v2, DeeperForensics-1.0, FaceForensics++, Korean Deepfake and Indonesia, totaling 3000 videos), the model achieved 92.54% accuracy, 91.69% precision, 93.33% recall, and 92.50% F1-score on the test set. These findings demonstrate that integrating dual-domain representation with data-efficient transformer architectures produces a robust deepfake detection system deployable in real-world scenarios where manipulation techniques continuously evolve.]]>
Copyrights © 2026
