learn

Detecting images generated by artificial intelligence has become a game of cat and mouse. And the worst thing is that it is going to get worse. For a time, we all began to focus on the hands and in the number of fingers that the AI ​​represented in the images of people through the diffusion mechanisms of the models. A few years ago it was obvious to see when an image was created by AI. Now, with image models and video increasingly precise, the task is much more complex. The good news is that there are still ways to detect if an image has been generated by AI, although seeing the pace at which the models advance, this may soon change again. Detecting them is less intuitive than before, but just pay attention to geometry, shadows and perspective. Basically, technical drawing. Who is behind this idea. Hany Farid, a specialist at the University of California at Berkeley and one of the world’s leading experts in image forensics, has spent more than two decades dedicated to determining whether a photo or video has been manipulated. Santiago Lyon, former director of photography for the Associated Press who now works in digital security at Adobe, describes Farid in a Science report as “a kind of dean of digital forensics”, precisely because he has been at it for so long. Farid helped found this discipline more than 20 years ago, and says that AI is the biggest challenge he has faced. Farid exemplifies his method with this image. If we draw a line towards the horizon between the tiles and the skirting boards, we see that the lines do not converge at a single point, which tells us that the image is generated by AI It’s hard to know what’s true and what’s not.. We are losing the ability to trust what we see. The combination of generative AI, capable of creating images almost indistinguishable from reality, and a warm regulation on social networks It makes the hoaxes amplify, making it increasingly difficult to know if what we are seeing is real or not. And in many cases, we don’t even care. Farid speaks directly of a “global war for truth”, with consequences for people, institutions and democracies. In a TED talk He said that he believes that the percentage of fake images on the Internet is close to 50%. It is no longer useful to focus on pixels. One of the first techniques Farid developed was based on the “noise” left by real cameras. An authentic photo is born from light hitting an electronic sensor; An AI image, on the other hand, emerges from a statistical process that converts random noise into an image consistent with the text requested. This very different origin left traces detectable at the pixel level. The problem is that generators have learned to imitate even those imperfections, sensor noise and lens artifacts. As explains Science report, many of Farid’s pioneering methods based on statistical relationships between pixels “no longer work well, if at all,” because AI images are created from scratch rather than edited over a previous photo. technical drawing. AI, says Farid, “doesn’t know physics, doesn’t know geometry, and does all kinds of atrocities.” And that’s where technical drawing comes in. According to Farid, these are the three fronts that we must examine: Vanishing points. In the real world, parallel lines (train tracks, floor tiles, the sides of a wall) converge toward a single point as they move further apart. It is a principle that artists have known for centuries, but that AI ignores because it does not understand three-dimensional space. If those lines don’t meet at a single point, the scene is physically impossible. Shades. The Sun is so far away that its rays reach the Earth practically parallel. That means that the lines connecting each object to the shadow it casts should also intersect at a point consistent with the position of the light. In many AI-generated images, those lines don’t even come close to crossing. Highlights. The same principle applies to mirrors, as lines connecting one point on an object to its reflection should converge at a vanishing point. When they don’t, the image is given away. The same thing happens in this image. If we draw a line that passes through both the vertices of each cube and the vertices of its projected shadow, we see that they do not converge at a single point either. Track accumulation. No technique is infallible on its own, and Farid insist in that the method consists of accumulating clues, as in an investigation. In his TED talk he exemplified this with an image made with AI of several soldiers looking forward. In it he detected the suspicious pattern in the noise, the absence of a coherent vanishing point on the walls and shadows that did not intersect. Three anomalies that gave clues that the image was not real. The underlying reason why this approach stands up better over time is that AI companies are not looking to fool forensic experts like Farid, but rather the average user, since we are at a much lower bar. As he himself says“the visual system forgives all kinds of nonsense in photos because it doesn’t care.” In this image, if we draw a line from a point in the figure to the same point reflected in the mirror, we see that the lines do not converge at a single point either. Doubts and limits. Not everyone in the field shares the same optimism. Some researchers reaffirm that each detection technique has a very short “useful life”, sometimes a few months, because AI improves very quickly. In fact, the famous mistakes on six-fingered hands disappeared in a flash. Farid, however, is skeptical that AI will ever master complex real-world physics, like an explosion, because simulating it is devilishly difficult and companies have little incentive to go that far. Still, he acknowledges that receives a dozen emails every day from journalists … Read more