Chrominance and Luminance: a study to detect deepfakes

A deepfake is a fast attack technique that has evolved rapidly over the past several years. It only takes one person to synthesize thousands of photorealistic images in a few hours or to manipulate a large number of videos. The creation of fake faces through image tampering has been identified as abusive media and could result in major ethical, legal, or political consequences. Existing known methods, such as Generative Adversarial Network (GAN), have simplified the synthetization of such images and can be used for detection purposes. However, current solutions and existing methods remain vulnerable and cannot stop deepfakes from spreading. In this paper, two novel solutions are proposed, named ColorDense and LightDense. They both use the DenseNet network backbone to determine fake face images from real face images. The two proposed models incorporate luminance and color signals from multi-color spaces to spot manipulated images obtained from the most recent deepfake datasets. Chrominance components of face images are considered to give a deep representation, while luminance depends on measuring the perceived brightness levels of the image. The system extracts the important features using the DenseNet network. Afterwards, it then passes the feature vector to a classification system based on neuronal techniques such as Multilayer Perceptron (MLP), Random Forest (RF), Support Vector Machine (SVM), and Radial Basis Function (RBF) as deepfake image detectors. Several experiments were conducted in this paper and the proposed system achieved over 98% accuracy in classifying deepfake images.