Read: “If we are not vigilant about Deepfake geography, then we will enter the danger of a ‘fake geography’ anti-utopia.”
Deepfake has a new “track” – faking satellite maps.
Deepfake refers to artificial intelligence technology that, after deep learning, produces fake images, such as “face swapping” videos of celebrities and politicians and producing pornographic videos with the faces of female celebrities. A paper published last week in the journal Cartography and Geographic Information Science suggests that deep forgery techniques have been able to create indistinguishable fake satellite maps, which may cause problems for people.
In an interview with The Verge, Bo Zhao, the first author of the aforementioned paper and an assistant professor of geography at the University of Washington, said, “While many GIS practitioners are appreciating AI technology and seeing its advantages in solving geographic problems, few will acknowledge or openly criticize AI’s deep falsification is a potential danger.”
In their research, Bo Zhao and team used empirical evidence to prove that deep falsification of satellite images is technically feasible. They selected satellite maps of three cities, Seattle and Tacoma in the U.S. and Beijing in China, and the satellite images of the three cities differed in green space, street width, etc. The artificial intelligence system will learn the urban landscape features of Seattle and Beijing, using Tacoma as the base dataset, and then generate a fake map.
The process of deep artificial intelligence falsification
The results showed that after 200 cycles of training, the AI system was able to successfully generate a satellite map of Tacoma that did not exist on Earth. Although the fake satellite map is not considered perfect, it has less color index than the real map and more complex and uneven textures, it still looks real and is difficult to recognize with the naked eye.
The researchers concluded that the deeply faked satellite map is more confusing than the usual images and videos because viewers can easily attribute unnatural areas on the map to the poor quality of the images and because the average satellite map does not need to be high-definition to be convincing. “Since most satellite images are generated by professionals or governments, the public is usually more willing to believe they are real.” Researcher Bo Zhao said.
a: Basic data map of Tacoma; b: Real satellite map of Tacoma; c and d are faked maps of Tacoma based on views of Seattle and Beijing, respectively. Image source: https://www.washington.edu/
When artificial intelligence learns to falsify satellite maps in depth, what can they be used for?
Researchers believe that there are many scenarios in which maps can be maliciously faked, such as scenes of mountain fires or floods, and that the existence of faked satellite maps makes real satellite maps less credible – when someone alleges that a map is faked, it is difficult for the average person to distinguish with the naked eye and determine who to believe. 2019 Back in 2019, analysts at the U.S. government-affiliated National Geospatial-Intelligence Agency also noted that deeply falsified satellite maps could mislead the military’s strategic deployments, such as when military training is misled by fictitiously created landmarks that do not exist.
“As technology continues to evolve, this research aims to encourage a more comprehensive understanding of geographic data and information and to debunk the appearance of absolute reliability of satellite imagery.” Bo Zhao said the goal of his research is to raise public awareness of deep falsification of maps, and that people need to be aware of this risk before they can better deal with this new problem.
A satellite map
Facing the problem of deep falsification of maps, the most crucial thing is to have the coping technology to check the authenticity of the maps. The paper also points out that the analysis of false maps is technically a “binary result”, i.e., completely true or completely false, and it is technically impossible to pick out the false parts of a map that is a mixture of true and false. If we want to train a system that can “pick out fakes” on maps, we need to build a more comprehensive satellite image database and then train a targeted dataset containing various real and fake satellite image scenes.
In addition, due to the complexity of real-world landscape features and the diversity of deep forgery methods, it is almost impossible to have a universal map authenticity identification system.
Bo Zhao believes that geography researchers should start developing authentication systems that can be made available to the public, which are similar to the current fact-checking tools for online information. For the public, as with media literacy, geospatial data literacy is important, and people should understand the impact of false geographic information on society.
“If we are not alert to the problem of deeply falsified geography, then we will enter into the danger of a ‘fake geography’ anti-utopia.” The research team warned.
Recent Comments