TY - THES ID - 134688297 TI - Morphed face generation using GANs AU - Maarten, Craeynest AU - Joosen, Wouter AU - Preuveneers, Davy AU - KU Leuven. Faculteit Ingenieurswetenschappen. Opleiding Master in de ingenieurswetenschappen. Computerwetenschappen (Leuven) PY - 2019 PB - Leuven KU Leuven. Faculteit Ingenieurswetenschappen DB - UniCat UR - https://www.unicat.be/uniCat?func=search&query=sysid:134688297 AB - Automatic border control systems have recently been shown to be vulnerable to a morphing attack. It attacks the authentication process by creating face images that can be used to match multiple people. This can then be used to create a passport that is valid for two or more individuals. Current methods to generate these morphs have been shown to have several problems: They are (1) very easy to identify using specialized detectors and (2) only work when there is a high similarity between the two input faces. Our goal is to find a new generation method that does not suffer from these shortcomings. To do this we explore generating morphs using generative adversarial networks. First we create an overview of the conditional neural network designs used for image generation. For each we discuss the core idea behind it and its advantages and disadvantages. We then work out two of these designs for the application of morph generation. Finally we compare their performance to the current state of the art of morph generation. To do so we introduce a new evaluation methodology that tests both the quality and morph similarity with human and automatic evaluation methods. Our approach achieves the same automated match rate of 60% compared to existing techniques, but results in morphs that are visually a better split between the inputs. The downside however is that our method has more artifacting. We conclude that generative adversarial neural networks show potential for morph generation, but miss a good way to teach the network about the similarity between two faces, which directly results in poorer quality morphs. ER -