Three-objective evolutionary search of the latent space of Generative Adversarial Networks for human face generation

This article presents multiobjective evolutionary approaches for searching the latent space of Generative Adversarial Networks for human face generation, considering similarity, gender, and race attributes. The problem is relevant to address and improve the robustness of automatic recognition systems. Three variants of Evolutionary Strategy optimization are proposed and compared using multiobjective optimization metrics, quality of generated images, and capability of deceiving the recognition system. Results indicate that the enhanced version of the hybrid evolutionary strategy using an adaptive grid to keep and promote diversity computes the most accurate solutions and proper coverage of the Pareto front. The outcomes of the research hold practical significance in enhancing the resilience of facial recognition systems.