Performance of Hybrid Models with VGG16-XGBoost and VGG16-LightGBM on Fake Face Image Recognition

We implement this research to evaluate the performance of the hybrid models on classifying real and fake images. The hybrid models theoretically have the same architecture of combination of two stages. The first stage is image feature extracting process and the second stage is image classifying process. In first phase, we employ a pre-trained CNN model, namely VGG16 to extract features of real/fake images. In the second phase, our hybrid model uses two advanced algorithms, namely XGBoost and LightGBM to make classification of images. Thus, the two hybrid models are named as VGG16-XGBoost and VGG16-LightGBM. The empirical implementation on both of datasets on fake face images, which were collected and produced by Flick social network, by experts in photo-shop and automated by GAN. The results of experiments show that the hybrid models given by VGG16-LightGBM produced higher accuracy than VGG16-XGBoost on the problem of classification on real and fake images.