Deep learning and AI have, in recent years, made us increasingly have to doubt the content we see. It is relatively affordable for anyone to replace someone else’s face in a photo, or even a video, as well as to generate new faces of people who do not exist. Now, they have managed to do that, but with images taken by satellite.
Deepfakes continue to fill the web, even being used to undress clothed people, or to generate faces of people who do not exist. The https://thispersondoesnotexist.com/ website has improved in recent years, offering increasingly real faces. Now, they are seeing the dangers of images of cities and countryside that can be modified to show content that is not really there.
Deepfakes hit the maps
Geographers are concerned about the proliferation of this type of content. For example, they can be used to create hoaxes related to fires or floods, or to say that a country has a concentration camp when in fact it does not. The United States Army already warned about this in 2019, where a military software can see a bridge in the wrong location, and that can affect the development of a mission, since they can trap them by not having a bridge in that location. .
A group of researchers has now published an article where they analyze this problem, called « Deep fake geography? When geospatial data encounter Artificial Intelligence «. This research is the first to apply the concept of deepfakes to satellite images, and they have shown that it can be done. With this, they want to inform citizens of their possible problems.
Researchers claim that humans have been lying for years with maps , including boundaries and margins where they were not, hiding islands, putting towns that do not exist or streets that are not really there. This was used to identify those who copied maps, and although in Google Maps we are not going to find things like that, in alternatives to Maps we can find content falsifications if the company is controlled by the government, as can happen in China.
Your AI can detect them, but they will get better
AI is now used in maps to solve geographic problems or identify features, but few people seem to have realized the dangers it entails. Researchers have managed to generate sites from scratch that look real, and someone who does not look in detail may think that they are real sites. For this, they have used the Generated Adversarial Networks , or GANs , which use other systems such as the web that generates the faces of people that do not exist.
To train him, they used real images taken by satellite of places like cities. The training process was not easy, as some of the images generated had some strange shadows.
The low resolution of satellite photos helps them to be faked relatively easily. With this detection system they could also detect false images by analyzing details such as texture, contrast or color. The problem is that these systems need to be constantly updated to compensate for the constant improvements and innovations that are made in the generation of deepfakes.