The use of facial biometrics in digital applications raises privacy concerns, especially under the PDP Law No. 27 of 2022, as the original image is at risk of being misused, accessed
without permission, to privacy violations and legal implications in forensic and research contexts. To e!ectively these issues, this study proposes a privacy-preserving methodology
that generates synthetic facial images using Deep Convolutional Generative Adversarial Networks (DCGAN) to maintain the utility of biometric data for gender classification
tasks while protecting individuals’ personal identities.
We trained the DCGAN model on anonymized facial image datasets obtained from real student facial data collection, using an adversarial learning framework designed to balance
image realism and privacy protection. As a comparative baseline, an f-divergence GAN (FGAN) model was also implemented and evaluated. The utility of the synthetic facial
data was assessed via gender classification using a fine-tuned Vision Transformer, while forensic detectability was evaluated on 1000 images using six techniques: Noise Analysis,
PCA, Metadata Extraction, Clone Detection, ELA, and JPEG Analysis.
Our experimental results demonstrate that DCGAN surpasses FGAN in generating high-quality, realistic synthetic facial images, with a notably low generator loss (0.3 compared to 0.9) and high discriminator accuracy (0.9 compared to 0.1). Gender classification performed on these synthetic images achieved an accuracy of 85.00% using GAN synthetic
data compared to 99.50% on real data, based on a test of 1000 synthetic and 1000 real images, indicating that the synthetic data still retains essential biometric cues needed for
analytic tasks. A forensic tool was used to assess the generated images, and the analysis identified synthetic generation traces in 99.97% of images, confirming their artificial origin
while retaining realistic facial features. This study developed a DCGAN-based method approach for generate synthetic facial data that preserves privacy while retaining utility
for gender classification and forensic analysis.