|
\justifying Recent advances in GAN technique have made it much easier than ever to generate believable face images, which brings some potential security issues to the public. Currently a variety of architectures have been proposed to detect these generated fake images, but few works address the problem of generalization ability of forensics models, which means most of them need prior knowledge of the structure of GAN model when detecting a specific GAN fake face images, and unable to detect unseen fake images generated by other GAN models. In this work, we tackle the generality enhanced problem by adding a fixed weight convolutional layer before CNN structure, which contain three designed high-pass filters. Firstly, we convert RGB images to YCbCr color space to get more obvious edge information. Then, EfficientNet is adopted as our basic CNN structure to extract inhenrent features and classify them. Finally, a fixed weight layer is added to the first convolutional layer, which could greatly suppress the semantic image content and magnify microscopic characteristics of images, thus help to enhance the generalization ability of the model. A series of experiments demonstrate that our approach achieve superior performance compare with the state-of-the-art work in terms of the generality of forensics model. |
|
Keywords:blind forensics, fake face detection, generability, deep learning, generative adversarial networks |
|