Assessing YOLOv8 as a Classifier for Detection of Synthetically Generated Facial Imagery

The literature has shown that facial recognition technology has been widely employed in diverse industries, such as consumer and security applications. It played a crucial role in enhancing consumer security measures, facilitating biometric identification processes, and elevating consumer experiences. However, the deployment of this technology also leads to significant privacy and ethical considerations that necessitate meticulous attention and regulation in operations. In recent years, You Only Look Once (YOLO) has also shown promising results as an image classifier. However, an in-depth analysis is required to explore the effectiveness of YOLOv8 in the classification of real and fake facial images. Hence, this research is to build an image classification framework employing the YOLOv8 architecture with the aim of efficiently differentiating between real and fake or synthetic (artificially generated) facial images. The methodology involves training the YOLOv8 model using a publicly accessible dataset containing both real and fake faces generated using generative artificial intelligence (AI). The dataset has been obtained from the Kaggle database, consisting of 140k real and fake images employing AI. The extensive series of experimental findings validates the efficacy of the model, with a notable training and testing accuracy of 99.35% and 98.2%, respectively. These significant results are obtained with training over 20 epochs with a minimal error rate of 0.0065. The aforementioned results are the best among standalone deep learning models trained using transfer learning, highlighting the practical efficiency and resilience of the proposed framework for discerning authentic and AI generated facial images, hence demonstrating its potential for use in diverse practical contexts.