AI Generated Face Detection Using DenseNet - 121

Abstract

In this digital era, AI-generated faces have become so realistic that distinguishing real from fake has become almost impossible. To address this issue, the findings presented in this paper were utilized to develop an advanced digital detection mechanism. The paper presents a DenseNet-121-based neural network trained on 65,000 real and synthetic facial images. The paper implemented the detection into a practical, user-friendly web application based on Flask and React, where users can upload their images, do necessary cropping, and instantly receive predictions along with confidence scores, essentially democratizing deepfake detection to the masses. When evaluating on 13,000 entirely new sets of images, it performed 76 errors (65 real faces wrongly flagged as fake, 11 fakes wrongly accepted), attaining 99.42% accuracy in just seven training epochs by using techniques such as data augmentation and check pointing, including image augmentation and adaptive check pointing. These results confirm the model’s effectiveness in distinguishing between genuine and AI-generated faces. The paper then concludes with future directions, including other video deepfake detection and ethical considerations concerning this technology, asserting that without enhanced digital literacy and public awareness programs, building safer and more trustworthy digital worlds would only be possible by having effective detection tools.