Unmasking deepfakes based on deep learning and noise residuals

Abstract EN

The main reason for the emergence of a deepfake (deep learning and fake) term is the evolution in artificial intelligence techniques, especially deep learning.

Deep learning algorithms, which auto-solve problems when giving large sets of data, are used to swap faces in digital media to create fake media with a realistic appearance.

To increase the accuracy of distinguishing a real video from fake one, a new model has been developed based on deep learning and noise residuals.

By using Steganalysis Rich Model (SRM) filters, we can gather a low-level noise map that is used as input to a light Convolution neural network (CNN) to classify a real face from fake one.

The results of our work show that the training accuracy of the CNN model can be significantly enhanced by using noise residuals instead of RGB pixels.

Compared to alternative methods, the advantages of our method include higher detection accuracy, lowest training time, a fewer number of layers and parameters.