Fake Faces Detection Based on an Auto-Encoder Network (Application Developed on Jetson Nano)

Abstract

Synthetic face generation has overgrown over the past few years, creating hyper-realistic forged images and videos based on Deepfake techniques. In this master's thesis, we propose a system that aims to detect fake faces in images based on a specific type of neural network called Convolutional Auto-Encoder (CAE). This unsupervised deep learning approach allows us to obtain a simple representation of our data on which we then use the estimate of kernel density and reconstruction error to distinguish between "false" and "real" images. ". We have opted for the use of the Convolutional Auto-Encoder known for its performance in detecting anomalies. We trained the AE on 50,000 samples from the Flickr database. The tests performed on the false faces StyleGan dataset yielded favorable results but can be refined by prolonging the training. The application is embedded in the NVIDIA Jetson Nano 2GB Developer Kit, considered a very powerful mini-computer because it has a GPU with 128 cores, specially designed for AI and robotics applications.