Citizen-based Monitoring for Peace & Security in the Era of Synthetic Media and Deepfakes
Speakers: Alexander Glaser, Princeton University & Vy Nguyen, BHT Berlin
Satellites have been used since the 1950s for Earth observation, first by governments for reconnaissance missions shrouded in secrecy, but increasingly also by a community of open-source analysts, researchers, and non-governmental organizations, enabled by a rapidly growing fleet of satellites operated by commercial providers. These collection efforts are creating an immense archive of digital data, and there is a widely shared expectation---or hope---that broad
and open access to such imagery will enable the timely detection of non-compliance with relevant international agreements. Here, we are particularly interested in monitoring compliance with nuclear nonproliferation and arms-control agreements, but similar opportunities are also emerging in the context of environmental and carbon-emission monitoring, emergency response and human-rights monitoring, and archaeological-site monitoring. Working with real satellite imagery has strong limitations, especially when used for research purposes. Access to high-resolution imagery can be extremely expensive when required at scale, and training change-detection algorithms is difficult when based on rather limited datasets.
As part of this project funded by the German Foundation for Peace Research, we seek to assess the long-term potential of satellite imagery for monitoring and verification purposes, leveraging synthetic satellite imagery generated with advanced machine-learning techniques. We show that large text-to-image models like DALL-E 2 or Stable Diffusion can achieve remarkable results in generating realistic high-resolution imagery with tunable characteristics. A major part of this effort also examines the growing risks and challenges that synthetic media pose for society and policy and seeks to help assess the viability of efforts aimed at confirming the authenticity of digital media and, in particular, the provenance and authenticity of satellite imagery.