Inception-powered Synthetic Image Detection based on Mesoscopic Neural Network Approach. Based on Afchar's apublication "MesoNet: a Compact Facial Video Forgery Detection Network".
Detailed tutorial, "Seeing is Believing — Mesoscopic Neural Networks for Synthetic Image Detection", available on GradientCrescent.
The are twoIpython Notebooks in this repository, covering the Meso4 and Meso4-Inception architectures, as well as the InceptionV3 architectures. These two notebooks are completely independent. Note that the InceptionV3 model has been commented out in both books - to run a benchmark, uncomment these and comment out the mesoscopic architecture instead.
500 images of FFHQ and StyleGAN datasets, representing true and synthetic data classes, respectively, were used to evaluate our models. Data is temporarily stored on a Google Drive, but is also available on the NVIDIA StyleGAN repository