machine learning – Is deepfake detection viable?

I’m thinking of doing a project on deepfake detection, but I’m not entirely sure if it is viable. Based on my understanding, how it works is that deepfake generation programs have a generative and discriminative network, and eventually after training, the systems reaches an equilibrium where the discriminative network can’t detect real vs fake faces. I was thinking of building a CNN-LSTM architecture where I analyze not only single image frames, but image frames over time as well to better discern between real videos and deepfakes, but I’m not sure if this is viable? Any help or resources would be appreciated.