Enhancing Privacy and Accuracy in Facial Recognition with Synthetic Data Generated by Pre-Trained GANs

Facial recognition systems have become widely integrated into a multitude of security systems. Their rapid spread has brought up some serious privacy worries due to the fact that personal information specifically faces are pulled from social media and is often used without consent. Due to this, a lot of these facial recognition systems tend to be demographically biased. It means they will not work equally well for all individuals; this research looks into the potential usability of synthetic data generated by StyleGAN2-ADA model pretrained with the CelebA-HQ dataset. This dataset would also include demographic attributes such as age, gender, and race assigned by the FairFace model to create a balanced dataset while preserving privacy. The synthetic dataset was tested through the ArcFace facial recognition model. The results when compared to models trained on real data, more specifically the CelebA dataset. Along with hybrid data, which is a mix of real and synthetic data. Metrics like accuracy, F-1 scores, and false positive rates (FPR) were considered. The results revealed that synthetic dataset achieved performance metrics comparable or in some cases better than real and hybrid datasets. This shows that synthetic data can effectively substitute real images for facial recognition systems. Though, it wasn’t true with Fairness Discrepancy Rates (FDR), which measured demographic fairness. FDR showed that synthetic datasets produced the lowest rates compared to real and hybrid data. The results show that while synthetic data offers privacy there are still some issues with fairness. This praxis indicates that synthetic data has potential as a tool for enhancing privacy protection in the realm of surveillance and cybersecurity, though at the same time, it still shows that improvement is needed so that it can wear all demographics. Note: This is the public dissemination copy of the doctoral praxis submitted to The George Washington University in partial fulfillment of the requirements for the degree of Doctor of Engineering. The official archival version, including accreditation and front matter, is available through ProQuest and GW ScholarSpace.