Deepfake Video Detection with Convolutional and Recurrent Networks

Conclusion and Future Work

Although deep learning methods are used to create deepfakes, deep learning methods can also be used to detect deepfakes. As demonstrated by the performance of the ensemble method previously described, deepfakes can be detected with high accuracy, helping to neutralize the negative effects of deepfakes. With future work, given additional computational resources, we could further improve the accuracy of our models by developing deeper CNNs and reducing the effect of vanishing gradients utilizing skip connections. Another idea worth exploring to boost accuracy is to add more models to the ensemble, training these models in parallel, and placing the output of each model in a feature vector format, which would then be passed to another model, trained to predict whether a video is real or fake based on the predictions of all models in the ensemble. Currently, our ensemble simply averages the output from the 3D CNN and CNN-LSTM to determine a final prediction.