TrustFoto Project

My current research project investigates how to distinguish fake photos from trustworthy photos. After all, you want to make sure the photos are authentic (i.e. without tampering) before publishing.

Some famous examples include Simpson's photo on Time Magazine, a copy-and-paste Iraqi war photo on LA Times, a shifted pyramid on National Geographics, and the photo of John Kerry and Jane Fonda on LA Times. These fake photos can be very misleading and easily impact a person's reputation.

I am now looking at how to solve this problem using a generative model approach. Basically, if we can tell there is some inconsistency in the photo, then we say the photo is suspicious.

Inconsistency includes semantics, illumination, acquisition process...etc. My current focus is the last one: acquisition process. If we can successfully determine there are two parts in the photo that are taken from two different cameras, then we claim this photo is suspicious. Therefore, a good camera model and a reliable camera calibration algorithm are our target.

For more information, please refer to the project page.