Generative Adversarial Networks for Anonymous Acneic Face Dataset Generation

The paper presents a method to generate a realistic anonymous synthetic dataset of human faces with acne at varying severity levels using a StyleGAN-based algorithm. This approach addresses the lack of large, balanced, and representative datasets in biomedical applications, particularly for pathological human face images. By leveraging generative adversarial networks, the method ensures privacy and legal compliance, as no real subjects are involved. The generated synthetic dataset was used to train a CNN-based classification system, achieving an accuracy of 97.6% with InceptionResNetv2 when tested on authentic face images. This work facilitates the use of synthetic datasets in scientific research without legal or ethical restrictions and can be extended to other medical imaging applications.