Frequency-Aware Explainable Deepfake Video Detection Using Dual-Branch Deep Neural Networks: A Review

Abstract

With the rapid development of autoencoder-, GAN-, and diffusion-based generative approaches, the artifacts resulting from deepfake video synthesis have reached a level that makes them nearly visually indistinguishable from authentic video, posing risks to society via misinformation and enabling identity theft or fraud. A recent, growing consensus is that any single visual feature alone is insufficient for detection of realistic synthetics; some purely spatial artifacts like blending seams and inconsistent pose might already have been skillfully removed during synthesis, while other frequency-domain traces stemming from, e.g., upsampling operators or decoders, remain, even when any pixel-level discrepancy is invisible. This led to the prevalent tendency towards dual or multi-branch networks to fuse spatial (RGB) data with frequency (FFT) data, often by harmoniously coordinating branches with cross-attention or adaptive-fusion modules. In parallel to this development, there is a significant and increasing enthusiasm for explainable AI (XAI) techniques, primarily Grad-CAM-like methods, which produce visually appealing rationales and contribute to making these detections accountable. This study presents a literature review on the most important works at the intersection of frequency-awareness, dual-branch fusion, and explainability in deepfake video detection. It establishes a taxonomy comprising pure spatial, pure frequency, and hybrid fusion methods; reviews important architectures (e.g., MesoNet, F3-Net, FreqNet-style dual branches, and fusion through cross-attention); compares them with values reported on FaceForensics++ and Celeb-DF (V2); reviews existing techniques to explain such detectors; and discusses remaining challenges. These challenges, such as generalization to new manipulation methods, sensitivity of frequency-information routing within the branches, overall computational efficiency, and the lack of XAI tailored to the dual-branch fusion mechanism itself, point toward directions for future work aimed at more integrated, intrinsically interpretable, frequency-aware dual-branch models.