Data synthesis using generative models


  • Fredrik Johansson
  • Andreas Horndahl
  • Harald Stiff
  • Marianela Garcia Lozano

Publish date: 2020-11-27

Report number: FOI-R--5041--SE

Pages: 80

Written in: English


  • Generative adversarial networks
  • generative models
  • deep learning
  • deep neural networks
  • synthesis


Within the research field of generative modeling, the possibilities for what can be achieved in terms of computer-based data synthesis are investigated. Deep learning-based generative models make it possible to learn the distribution of a large unannotated training dataset and use this for generating new synthetic data of good quality that is hard to tell apart from real data. The development of generative modeling is fast-paced, and new methods and techniques for synthesizing and manipulating high-dimensional data such as facial images, news articles, voices, or videos are constantly appearing. The aim of this report is to provide an understanding of how generative modeling creates threats as well as opportunities for military-related applications. The conducted literature survey indicates that there are several possibilities of using generative models in military applications, e.g., for improving the realism of military exercises and facilitating communication. A lack of powerful hardware and the limited availability of representative data are identified as potential barriers of using generative models in practical applications.