Great initiative, Abhishek. One angle worth baking in early: the most robust signal usually is not in perception (blinking, lip-sync) but in the frequency domain. GANs and diffusion leave a periodic upsampling fingerprint in the spectrum that a 2D FFT plus a radial average collapses into a single score. Pairing that with Error Level Analysis for splicing and a face-landmark check (MediaPipe) gives you an ensemble that degrades gracefully instead of one brittle model. I put together a Python walkthrough of exactly this ensemble over on Meu Universo Nerd if it helps your build.