Abstract:
In the last few years, the explosive development of Generative AI technology, particularly Generative Adversarial Networks (GANs), has transformed the digital media construction world, whereby ultra-realistic images can now be produced that closely resemble human images. These are the AI-produced images, or the so-called deepfakes that are so realistic that even trained specialists have a hard time distinguishing the real from them. Although those innovations have catalysed creative potential in all sorts of industries like film, gaming, and advertising, they have also been accompanied by serious fears about digital trust, authenticity and the propagation of misinformation.
The increasing presence of synthetic media contents, especially AI faces, is causing serious concerns in such fields as journalism, law enforcement, national security, and public discourse. The simplicity by which such images could be fabricated and spread represents a flaw within the visual proof and can alter mass consciousness on a wholesale scale. Against these challenges, this project aims at developing, constructing, and deploying a practical AI-based image classification system that can reliably tell the difference between real and AI images using deep learning-based methods.
In order to do that, a supervised machine learning model was built and trained with a publicly available dataset found in Kaggle and called the Real and Fake Faces dataset. This dataset is images that are labelled and either assigned “real” or “fake”. It is a good base for binary classification. The complete machine learning pipeline had been built, i.e., from dataset management to the training and optimization of the model, using TensorFlow and Keras in a Google Colab environment. The architecture used on this task was derived from MobileNetV2, a small and efficient Convolutional Neural Network (CNN) that showed excellent performance in mobile and embedded applications. MobileNetV2 has been selected because of its speed-to-accuracy trade-off and its compatibility with TensorFlow Lite for mobile deployment.
