Deepfake Technology Applied to Satellite Maps

The Emerging Threat of AI-Generated Synthetic Imagery

The discussion surrounding deepfakes, or artificially generated imagery, has largely focused on non-consensual depictions of individuals. However, this technology presents broader dangers and intriguing possibilities. Recent research demonstrates its potential to manipulate satellite imagery, creating convincingly realistic, yet entirely fabricated, overhead maps of urban areas.

Research into Manipulating Cartography

Led by Bo Zhao at the University of Washington, this study isn't intended to incite alarm. Instead, it aims to highlight both the risks and the potential benefits of applying this controversial technology to the field of mapmaking. The team’s methodology shares similarities with style transfer techniques—altering images to resemble impressionistic paintings or drawings—rather than traditional deepfake applications.

Training the Machine Learning System

The researchers trained a machine learning system using satellite images from three distinct cities: Seattle, Tacoma (located near Seattle), and Beijing. Each city possesses a unique visual character, akin to the style of a particular artist.

For example, Seattle is characterized by abundant overhanging vegetation and relatively narrow streets. Conversely, Beijing appears more monochromatic in the images used, with taller buildings casting prominent shadows. The system learned to correlate details from street maps (such as those provided by Google or Apple) with corresponding features in the satellite views.

Generating Faux Satellite Imagery

When presented with a street map, the resulting machine learning agent generates a realistic-looking, artificial satellite image depicting how that area would appear if it were located in any of the three cities. The image provided illustrates this capability, with the top right image representing Tacoma, while the lower images showcase potential appearances in Seattle and Beijing.

Image Credits: Zhao et al.

Limitations and Plausibility

Upon closer examination, the generated maps exhibit a slight lack of sharpness compared to authentic images. Logical inconsistencies, such as dead-end streets, may also be present. Nevertheless, the Seattle and Beijing images are remarkably plausible at first glance.

Potential Applications – Both Legitimate and Otherwise

Numerous applications, both constructive and potentially harmful, can be envisioned for this technology. The researchers propose its use in simulating imagery for locations lacking existing satellite data—such as cities before the advent of satellite technology, or areas undergoing planned development or zoning changes.

The system isn’t limited to imitating entirely different locations; it could be trained on denser areas within the same city or regions with wider streets.

Reviving Ancient Maps

It could even be employed, as demonstrated in a related project, to create realistic modern maps from historical, hand-drawn ones.

Detecting Simulated Imagery

The paper also explores methods for detecting such simulated imagery, focusing on meticulous analysis of colors and features, should the technology be misused.

Challenging Assumptions About Geospatial Data

According to Zhao, as reported in a UW news article, this work challenges the common belief in the “absolute reliability of satellite images or other geospatial data.” Similar to other media types, this assumption must be re-evaluated as new threats emerge. The complete research paper is available in the journal Cartography and Geographic Information Science.