Fake Image Detection Using Deep Learning

Abstract

With the emergence of numerous electronic communication programs and image processing programs, as well as an increase in the number of people who use them with a zeal for publishing everything related to their lives and their special pictures and their fear of those who might use these pictures for malicious or humorous purposes, it has become necessary to have specialized and precise systems to determine whether a picture is real or fake. Our work aims to detect real and fake faces by using and modifying one of the most efficient CNN architectures, EfficientNetB0, after improving the architecture with additional fully connected layers and efficiently training the model by using the Adam optimizer and a scheduler learning rate technique. Our findings on the well-known 140k-real-and-fake-faces Kaggle dataset showed state-of-the-art accuracy with the lowest error rate. We achieved 99.06% accuracy, and 0.0569 error rate respectively.