Quad-Stream Deepfake Detection: Combining Spatial and Frequency Domain Analysis for Robust Video Authentication

Deepfake videos poses a great challenge in cyber-security, it may deceive people to take actions that increase the vulnerability to fraud, misinformation, identity theft, and social engineering attacks. Current methods, for the most part, take as input RGB images and process them as a blackbox algorithm that learns to distinguish between real and fake. In this paper, we propose a quad-stream deepfake detection network that combines spatial and frequency domain analysis of both the face region and the entire video frame. Our approach first performs face detection and alignment to extract high-quality face crops from video frames, while also retaining the context of the whole frame. Each face crop and full-frame image is then transformed into both RGB spatial domain and frequency domain representations (via Fourier transform), yielding four parallel input streams. We design a neural network architecture with four specialized CNN-based sub-networks: two spatial streams (face and frame) using pretrained backbones, and two frequency streams (face and frame) using modified ResNet backbones for frequency spectrum input. Features from all streams are combined for final classification. The proposed method shows promising results as it is able to spot subtle forgery artifacts in both facial details and broader scene context, as well as anomalies revealed in frequency spectra that are imperceptible in raw pixels. We evaluate our model on popular deepfake video benchmarks including FaceForensics++ and Celeb-DF v2, achieving state-of-the-art detection accuracy. Notably, our quad-stream approach improves robustness under challenging conditions such as video compression and cross-dataset evaluation. An ablation study confirms that each stream (face vs. frame, spatial vs. frequency) contributes to overall performance. Our results demonstrate that integrating spatial and frequency information at multiple scales yields a significantly more generalizable deepfake detector.