While the idea of “deepfakes,” or AI-generated artificial imagery, has been decried primarily in reference to involuntary depictions of individuals, the expertise is harmful (and fascinating) in different methods as effectively. For occasion, researchers have proven that it may be used to govern satellite imagery to supply real-looking — however completely faux — overhead maps of cities.
The examine, led by Bo Zhao from the University of Washington, isn’t supposed to alarm anybody however somewhat to indicate the dangers and alternatives concerned in making use of this somewhat notorious expertise to cartography. In reality their strategy has as a lot in frequent with “style transfer” methods — redrawing pictures in an impressionistic, crayon, and arbitrary different fashions — than with deepfakes as they’re generally understood.
The crew skilled a machine studying system on satellite pictures of three completely different cities: Seattle, close by Tacoma, and Beijing. Each has its personal distinctive look, simply as a painter or medium does. For occasion, Seattle tends to have bigger overhanging greenery and narrower streets, whereas Beijing is extra monochrome and — within the pictures used for the examine — the taller buildings solid lengthy, darkish shadows. The system discovered to affiliate particulars of a road map (like Google or Apple’s) with these of the satellite view.
The ensuing machine studying agent, when given a road map, returns a realistic-looking fake satellite picture of what that space would seem like if it have been in any of these cities. In the next picture, the map corresponds to the highest proper satellite picture of Tacoma, whereas the decrease variations present the way it may look in Seattle and Beijing.
An in depth inspection will present that the faux maps aren’t as sharp as the actual one, and there are in all probability some logical inconsistencies like streets that go nowhere and the like. But at a look the Seattle and Beijing pictures are completely believable.
One solely has to assume for a couple of minutes to conceive of makes use of for faux maps like this, each professional and in any other case. The researchers counsel that the method may very well be used to simulate imagery of locations for which no satellite imagery is offered — like considered one of these cities within the days earlier than such issues have been doable, or for a deliberate enlargement or zoning change. The system doesn’t should imitate one other place altogether — it may very well be skilled on a extra densely populated a part of the identical metropolis, or one with wider streets.
It might conceivably even be used, as this somewhat extra whimsical challenge was, to make realistic-looking trendy maps from historic hand-drawn ones.
And ought to expertise like this be bent to much less constructive functions, the paper additionally appears at methods to detect such simulated imagery utilizing cautious examination of colours and options.
The work challenges the overall assumption of the “absolute reliability of satellite images or other geospatial data,” stated Zhao in a UW news article, and positively as with different media that sort considering has to go by the wayside as new threats seem. You can learn the total paper on the journal Cartography and Geographic Information Science.