A Comparative Study of EfficientNetB4 and VGG19 Models for Deepfake Detection

With the increasing production of hyper-realistic altered images, the need for effective deepfake detection technology has become critical. These altered images pose significant threats to security, privacy, and the spread of misinformation, complicating the distinction between authentic and manipulated content. This challenge has far-reaching implications, from social media and politics to personal relationships. This study focuses on detecting deepfake human face images using a balanced dataset of 140,000 images, comprising 70,000 real faces sourced from Nvidia’s Flickr dataset and 70,000 fake faces generated by StyleGAN. In this research, we compare the performance of EfficientNetB4 and VGG19 models, to identify subtle manipulations in high-quality deepfake images. Our findings demonstrate that the EfficientNetB4 model achieves an accuracy of 98.54%, while the VGG19 model reaches 99.11% in the base model, highlighting the effectiveness of these models in advancing deepfake detection technology.