Empirical comparison of evolutionary approaches for searching the latent space of Generative Adversarial Networks for the human face generation problem
Jimena Mignaco, Gonzalo Rey, Jairo Correa, Sergio Nesmachnow, Jamal Toutouh · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2024
This article presents an empirical comparison of evolutionary search methods for exploring the latent space of Generative Adversarial Networks for the human face generation problem. Single- and multiobjective evolutionary algorithms are evaluated for the problem of optimizing two metrics that evaluates the similarity of generated images to a target one and the similarity to a target race attribute. The studied evolutionary algorithms are integrated into a software pipeline that also includes StyleGAN3 for human face images generation and DeepFace as the face recognition model applied to evaluate the similarity of generated images to the targets. The evolutionary algorithms explore the real-coded latent space of StyleGAN3, applying weighted sum and explicit Pareto dominance approaches. Most of the generated images are able to deceive the face recognition model in DeepFace, and numerical results demonstrate the superiority of the multiobjective approach regarding quality and diversity metrics. The obtained results have practical applications in improving the robustness of face recognition systems.