A Multi-Stream Model for Deepfake Face Detection Using Spatial, Frequency And Temporal Analysis

This project presents a multi-stream deepfake detection system designed to identify manipulated facial videos. To counter advanced generation methods, the architecture combines three analytical domains. Spatial features are extracted via a ResNet18 network, while hidden generative artifacts are uncovered using Fast Fourier Transform (FFT) for frequency-domain analysis. Furthermore, a Bi-LSTM network captures temporal inconsistencies across sequential video frames. These diverse features are integrated through an attention-based fusion module to maximize classification robustness. Validated on a processed video dataset, the system achieved a test accuracy of 94.21%, demonstrating significantly improved reliability over conventional single-stream detection approaches.