DeepFake Videos Detection using Hybrid Deep Learning Models
Kurella Tejasvi, Srisai Krishna Cherla, Uday Kiran Kairamkonda, G. Ramesh, Katam Pranathi, N. Siva Rama Krishna · 2025
DeepFake videos are artificial media produced through sophisticated artificial intelligence methods, frequently utilizing generative models. In order to solve this issue we came up with solution of identifying the DeepFake videos by combining ResNext and LSTM deep learning architectures. ResNext is employed for capturing facial patterns and extracting spatial features whereas LSTM inspects temporal changes frame sequences to identify manipulation over time and renders a verdict on the video as real/fake. Fake-face videos mislead viewers and speed up the spread of misinformation. Many detectors miss subtle artifacts when faces move fast or lighting shifts. We tackle this issue with a hybrid network that combines ResNeXt-50 for spatial cues with a two-layer LSTM that tracks frame-by-frame changes. The model learns from 19,800 clips drawn from FaceForensics and DFDC after face alignment and 10 Hz frame sampling. Training uses cross-entropy loss, batch size 32, and the Adam optimizer at 1×10⁻⁴. On a held-out DFDC test split the detector reaches 87.2% accuracy, 88.9% precision, and 86.0% recall while processing 22 frames each second. The hybrid stack lifts recall by six points over a ResNeXt-only baseline and keeps memory below 3 GB. These results show that fusing spatial and temporal signals in one compact model curbs the reach of manipulated videos on social platforms.