MIT Researchers Develop AI That Can Spot And Create Fake Images

Developing digital renders or editing them can eat up a lot of time. But, researchers from MIT and IBM might have a solution to this issue. The team has trained an AI to generate photographic images from edit objects in them.

While this could come in handy for artists and designers, it will also provide insight into how neural networks understands the context. The team aims to spot fake and morphed images. Dubbed GANpaint, the tool can be used as a part of the free demo. Instead of making manual changes, the tool will place an object accordingly to the scene.

The tool also lets you erase objects. It's still under development, but the team hopes the tool might soon be able to edit video clips. For instance, if the scene doesn't have a required prop, the editor could use the AI to add it later.

While building the AI, the team was surprised to discover that the system learned simple relationships among the object - such as the doors won't appear in the sky and others. While the tool makes it easier to create fake imagery, it could also come in handy to spot the fake images.

"You need to know your opponent before you can defend against it," said Jun-Yan Zhu, co-author of the paper explaining GANpaint. The researchers will present their work at a conference next month. In the meantime, you can give GANpaint Studios a spin.

Besides, a new study from New York University Tandon School of Engineering suggests that it would be really convenient if the process of picking fake pictures started from the source, the camera.

If the cameras become intelligent, it could help additional machine learning programs to identify if the image was morphed. The researchers also found out that the technique amplified a computer's accuracy at spotting fakes from 45% to 90%.

The method reimages the way a camera processes information to develop a picture. The camera creates unique artifacts in the image that software can later use to gauge whether or not the image was altered.