Deepfake Image Detection using Transfer Learning: A Comparative Analysis

Rapid advancements in the field of generative adversarial network (GAN) technologies have made the detection of AI-generated facial imagery extremely crucial. In this paper, we investigate the efficacy of transfer learning in distinguishing between real and AI-generated fake facial images. We used a comprehensive data set comprising 70,000 real face images from the Flickr data set and 70,000 synthetic faces sampled from a large-scale StyleGAN-generated data set. The combined data set was preprocessed by resizing all images to 256 pixels and partitioning them into training, validation, and testing sets. We evaluated the performance of several pre-trained deep learning architectures, including a Custom Convolutional Neural Network (CNN), MesoNet, ResNet50, EfficientNetB0, and Xception, through fine-tuning on our curated dataset. The results we achieved in this experiment show the importance of the use of, we managed to achieve accuracy of 98. 57% while using the Xception model in successfully distinguishing between real and fake faces. This study highlights the effectiveness of using established convolutional architectures for robust deep-fake detection.