Multi-target evolutionary latent space search of a generative adversarial network for human face generation Benjamín Machín Universidad de la República Uruguay benjamin.machin@fng.edu.uy Sergio Nesmachnow Universidad de la República Uruguay sergion@fng.edu.uy Jamal Toutouh ITIS Software, Universidad de Málaga Spain jamal@lcc.uma.es ABSTRACT This article presents an evolutionary approach for multi-target synthesized human face image generation based on exploring the latent space of generative adversarial networks. The proposed ap- proach seeks to generate diferent human face images those share similarities to two given target images. The optimization applies generative adversarial networks for face generation, facial recogni- tion for similarity evaluation, and an ad-hoc evolutionary algorithm for exploring the search space. The main results show that realis- tic images are generated, properly blending the main features of the two given target images, and deceiving a well-known facial recognition system. CCS CONCEPTS · Computing methodologies → Bio-inspired approaches; Neu- ral networks; Continuous space search. KEYWORDS generative adversarial networks, multiobjective optimization; evo- lutionary latent space exploration, human face image generation ACM Reference Format: Benjamín Machín, Sergio Nesmachnow, and Jamal Toutouh. 2022. Multi- target evolutionary latent space search of a generative adversarial network for human face generation. In Genetic and Evolutionary Computation Confer- ence Companion (GECCO ’22 Companion), July 9ś13, 2022, Boston, MA, USA. ACM, New York, NY, USA, 9 pages. https://doi.org/10.1145/3520304.3533992 1 INTRODUCTION Generative machine learning models are a class of statistical mod- els that are able to generate new data instances. Given a set of data instances  and a set of labels  , generative models try to learn/capture the joint probability  (, ) to describe the process of generating new datasets [6]. Generative Adversarial Networks (GANs) are deep learning-, neural network-based generative models [7] GANs consist of two artifcial neural networks (ANNs): a generator and a discriminator that apply adversarial learning to optimize their parameters. The generator  learns how to create synthesized samples  ′ from input Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for proft or commercial advantage and that copies bear this notice and the full citation on the frst page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specifc permission and/or a fee. Request permissions from permissions@acm.org. GECCO ’22 Companion, July 9ś13, 2022, Boston, MA, USA © 2022 Association for Computing Machinery. ACM ISBN 978-1-4503-9268-6/22/07. . . $15.00 https://doi.org/10.1145/3520304.3533992 vectors drawn from a random latent space  , i.e.,  ( ) =  ′ . The goal of the generator is to approximate the true data distribution. Simultaneously, the discriminator learns to distinguish real samples  (from a training dataset) from synthesized samples  ′ , produced by the generator. The GAN training process works iteratively and converges to a generator that approximates the real data distribu- tion and generate images that deceive the discriminator, which is no longer able to diferentiate real from synthesized samples (it labels real and synthesized samples at random). GANs have positioned as a successful tool for many applications that require the creation of creating synthesized data, especially those concerning multimedia (e.g., images, sound, video), healthcare, and other areas [13, 17ś19]. The synthesized data created by the generator is determined by the input vector drawn from the latent space. The latent space in a GAN is defned by a high dimensional random distribution (e.g., Gaussian distribution) linked to the trained generative model. Properly searching the latent space can lead to determining useful directions for generating images with specifc features, or fnd- ing disentangling sub-spaces or hyperplanes of the search space that control pose and facial characteristics. Given that the search space defned by the latent space of GANs has a large dimension, traditional optimization/search methods are not applicable. Thus, population-based metaheuristic methods, and especially evolution- ary computation [11], are useful to perform the search in reduced computational times. This article presents a proposal for fnding a sub-latent space (i.e., vectors from the latent space) that produces synthesized data samples similar to a set of target images. GANs are used to produce realistic human face images, and the search focuses on fnding latent space vectors to generate human face images with specifc attributes of the input images. This article extends our previous research [9], where an evolu- tionary algorithm (EA) was applied to search the latent space of GANs. The methodology applied in this article considers a multi- target approach, where the evolutionary search seeks to generate images with the features of two target faces. The generated im- ages are realistic, demonstrating that the proposed EA is capable of mixing features of the two input target images. Furthermore, the generated images are not distinguishable from the two target images for FaceNet, a widely used facial recognition systems [14]. Thus, this article studies the capability of generative neural mod- els to produce samples from the whole data distribution for the human face image generation problem and the efectiveness of evolutionary approaches for properly searching the latent space of GANs, when considering multiple target images. The main contri- butions of the research reported in this article are: i) a method based on EAs to generate synthesized human face images that properly 1878