A Systematic Review on Detection of Fake Video Through Deep Learning

Generative models such as GANs and diffusion systems enable the creation of highly realistic fake videos, eroding confidence in online content across social, political, and legal domains. Synthesizing insights from more than 85 scholarly articles published between 2018 and 2025, this work categorizes detection methods into spatial CNNs that identify frame-level flaws, temporal RNNs/LSTMs for motion inconsistencies, RPG-based physiological cues, and fused audio-video approaches. Evaluations on datasets like Celebs and Wild Deepfake yield accuracies above 95% in controlled settings, but cross-dataset generalization and defenses against advanced forgeries falter. Hybrid architectures with transformers emerge as leaders, revealing critical gaps in real-time efficiency and edge-device applicability to steer forthcoming innovations.