But the main type of neural network used in deepfakes is the “autoencoder.” An autoencoder is a special type of deep learning algorithm that performs two tasks.