Detecting Deepfake images using deep convolutional neural network

Abstract

In recent years, advancement in the realm of machine learning has introduced a feature known as Deepfake pictures, which allows users to substitute a genuine face with a fake one that seems real. As a result, distinguishing between authentic and fraudulent pictures has become di cult. There have been several cases in recent years where Deepfake pictures have been used to defame famous leaders and even regular people. Furthermore, cases have been documented in which Deepfake yet realistic pictures were used to promote political discontent, blackmail, spread fake news, and even carry out false terrorism attacks. The objective of our model is to di erentiate between real and Deepfake images so that the above mentioned situations can be avoided. This project represents a deep CNN model with 13000 images divided in two segments: Training and Testing. The dataset was prepared using necessary image augmentation techniques. A total of 2 categories are considered (real image category and fake image category). For testing purpose we have used a total number of 3000 images divided into two parts for real and fake class, each consisting 1500 images. 75% of the whole data was used as Testing data and remaining 25% as Training data. The dataset was tested against a custom CNN model referred to in the paper as the 18-layered CNN model and ve of the transfer learning models. Our suggested model was successful in achieving 98.77% accuracy whereas the best result out of the transfer learning model was achieved by InceptionV3 with 97.10% testing accuracy. The custom CNN model shows promising results in the case of detecting real and DeepFake images than all the other models used before.