Abstract:

Fake face images and videos are the most common type of digital manipulation. These content which are usually produced for entertainment purposes can have harmful consequences. Machine learning algorithms have started to be used in recent applications in fake face image processing. With machine learning algorithms, realistic facial manipulations can be made. Therefore, it has become very difficult to distinguish between real and fake content. Face manipulations are divided into 4 basic groups; entire face synthesis, face identity manipulation (deepfake), facial attribute manipulation, and facial expression manipulation. Faces that are not real are produced using entire face synthesis generative adversarial networks. In the identity change method the face image of the person in the video is replaced with another face. Attribute manipulation of face can be changed by skin, gender, age, glasses, hair color, etc. changing its properties. Expression manipulation method is to change the facial expression of the person. In this study, lightweight convolutional neural network algorithms were used to detect fake face images produced by entire face synthesis manipulation method. MobileNet, MobileNetV2, EfficientNetB0 and NASNetMobile algorithms were used for the training process. The dataset used includes 70,000 real images in the FFHQ dataset and 70,000 fake images produced with StyleGAN2 using the FFHQ dataset. In the training process, the weights of the models trained on the ImageNet dataset were reused with transfer learning. In the EfficientNetB0 algorithm, the highest accuracy rate was achieved with a accuracy of 93.64%.

Keywords: