DEEP LEARNING-BASED SYSTEM FOR DETECTING EDITED AND ORIGINAL PORTRAIT IMAGES

Abstract

AI has enabled the creation of highly realistic synthetic portrait images, commonly known as "deepfakes," which pose significant risks to digital trust and identity security. This thesis presents a deep learning–based system designed to automatically detect algorithmically manipulated portrait images and distinguish them from original photographs. The proposed solution leverages transfer learning with state-of-the-art convolutional neural network architectures, including ResNet50, EfficientNet-B0, and Xception. A two-stage training strategy—frozen-backbone training followed by fine-tuning—is employed, and model performance is evaluated using area under the receiver operating characteristic curve (AUC). Experimental results demonstrate that the selected model achieves robust generalization on unseen test data. Furthermore, a Gradio-based demonstration system enables real-time analysis of portrait images. The findings confirm the effectiveness of deep learning approaches for detecting AI-mediated image alterations, contributing to efforts in digital security and the prevention of misinformation.