Is deepfake satellite imagery the next battleground?
As deepfake technology becomes more potent, experts warn that “deepfake geography” may be an increasingly pressing problem.
Globally, the alternative data market reached around $1.6 billion in 2020, with from Research and Markets estimating that it will reach $11.1 billion by 2026. Alternative data is useful because it falls outside of the more traditional data sources used by organizations.
Consultancy firm Deloitte highlights the value of alternative data, especially for investment management (IM) firms.
"In the near future, IM firms will likely use news feeds, social media, online communities, communications metadata, satellite imagery, and geospatial information—to name a few data sets—to augment their traditional processes for securities valuation as the rule, rather than the exception," the authors write.
For this data to be valuable, however, it needs to be accurate and reliable. Research from the University of Washington suggests that geospatial imagery might be the latest victim of deepfake attacks. The researchers highlight the rise in what they refer to as “location spoofing”, which is when geospatial images are faked to mislead people.
What’s more, as deepfake technology becomes more potent, they warn that “deepfake geography” may be an increasingly pressing problem. The researchers set out to try and construct reliable ways of detecting fake satellite images, and call for a reliable means of fact-checking geospatial images. They warn that this is beyond simply photoshopping images to resemble the real thing, as deepfake technology can look incredibly realistic.
Of course, deliberate inaccuracies have been a part of mapmaking from their earliest times, due in large part to the inherent difficulties in translating information from real life into a map form.
While most inaccuracies are simple and honest mistakes, so-called “paper towns” are deliberate insertions of features such as mountains, rivers, and even cities into a map to avoid copyright infringements.
One notorious example of this was the inclusion of the fictional cities of Goblu and Beatosu in the official highway map from the Michigan Department of Transportation in the 1970s. The cities were inserted as a joke because the head of the department wanted to promote his alma mater, so inserted two places that were a play on “Go Blue” and “Beat OSU” while simultaneously protecting the copyright of the map.
Digital information systems
These japes seem almost quaint in an era dominated by digital mapping systems, such as Google Maps and Google Earth, which make the task of location spoofing far harder. The widespread use of these tools also means that such manipulation can have grave consequences, as the National Geospatial Intelligence Agency highlighted in 2019.
The Washington researchers examined how easy and effective deep fakes could be in terms of satellite images. They used the same framework that had been used in the manipulation of other forms of imagery to the field of mapping. They found that the algorithm was able to learn the crucial characteristics of each satellite image, before then generating a fake image based on those learned characteristics into a new base map. The researchers explain that it’s a similar approach to when software imposes the face of a human onto a cat.
They then took maps and satellite images from Seattle, Beijing, and Tacoma, and combined them to allow for the features to be compared. This was then used to allow the algorithm to create new images on one of the cities based upon characteristics from the other two. Tacoma was selected as the base map, with the features of Seattle and Beijing then used to incorporate new features into the deepfake map of Tacoma.