Exploring Gender Classification Performance on AI-Generated Facial Datasets Using Transfer Learning Models

Abstract

The demand for gender classification has increased significantly with the growth of smart and automated security systems. Artificial Intelligence (AI), particularly Deep Learning (DL), has emerged as a promising approach for building reliable classification systems. Despite advancements in classification techniques, the availability of high-quality datasets for training and testing remains a notable challenge. In this work, we propose utilizing a synthetic facial dataset to train several well-known classification models. We evaluate different Convolutional Neural Network (CNN) architectures, including transfer learning approaches, on a set of unreal face images generated entirely using a Generative Adversarial Network (GAN). To the best of our knowledge, this is among the few works that investigate gender classification trained exclusively on fully synthetic GAN-generated datasets, highlighting its novelty. The models achieved strong performance on the synthetic dataset, with up to 99% training accuracy and 95% test accuracy. However, as a limitation, the generalization of these results to real-world datasets remains uncertain, since synthetic images may not capture all demographic and natural variations. This study demonstrates the viability of using artificially generated data when real data is scarce or difficult to obtain.