Fake face detection based on a multi discriminator deep CNN architecture (MDD-CNN)

Abstract

Due to the robustness of the deep learning tools used to design these applications, fakes are becoming increasingly common as these applications become more widely available and accessible to the general public. These fakes are typically fake faces or even fake people, which are difficult to distinguish from real individuals. Therefore, we need more efficient applications for fraud detection. In this work, we propose a new multi-discriminator architecture to distinguish fake faces from real ones. The architecture consists of three deep networks (discriminators) competing with each other, each trained differently. The final decision is made by voting based on the decisions of the three discriminators. The core element of our architecture is the proposed new adversarial deep network discriminator (NDGAN), which is trained in three different ways, resulting in three distinct discriminators. Discriminator 1 undergoes adversarial training, discriminator 2 is trained using transfer learning, and the third discriminator undergoes supervised training with a standard CNN using examples and counterexamples. Training and testing were performed on 70 000 real faces from the Flickr-Face-HQ (FFHQ) dataset, while 70 000 fake faces were generated using Nvidia’s StyleGAN. The tests conducted on the three networks produced significant results, with accuracy ranging from 79 % to 98 % for fake faces, and from 80 % to 98 % for real faces. The reliability of the discriminators contributes significantly to the overall performance of the multi-discriminator system, achieving an accuracy of 96 % for fake faces and 98 % for real faces.