How to detect altered or AI-generated images

Home " How to detect altered or AI-generated images

Most of us like to make and see beautiful images. But it's hard to "make" a nice picture all at once without modification". photo-shop" . Except with artificial intelligence.

But how do you spot these changes? With 2 deep learning algorithms, it's possible.

PSCCNET: Identify modifications to an image

This algorithm (PSCC-Net: Progressive Spatio-Channel Correlation Network for Image Manipulation Detection and Localization) is ideal for detecting modifications to images. It can detect copy-paste, add, modify and delete.

Let's take an example:

Left: Unmodified image - Right: Modified image with background change

Left: No reaction, little or no change - Right: Strong reaction, sign of obvious changes — Left: No reaction = little or no change - Right: Strong reaction = clear change

TRUFOR: Spotting images generated by artificial intelligence (AI)

This algorithm (TruFor: Leveraging all-round clues for trustworthy image forgery detection and localization) to detect modifications, and is particularly effective in detecting images generated by artificial intelligence.

Let's take an example:

Left: An unmodified image with the cat's head, pasted from another image - Right: An AI-generated image — Left: An unmodified image, except for the cat's head, pasted from another image - Right: An AI-generated image

Left - Little reaction except on the cat's head, proving a modification - Right, a fairly global reaction proving that the image was generated by AI — Left - Little reaction apart from on the cat's head, proving a modification - Right, a fairly global reaction proving that the image has been generated by AI

Conclusion: Can we really detect images that have been modified or generated by artificial intelligence?

Of course, these algorithms are not perfect and there will be always false positives and true negatives. On the other hand, it's interesting to see that, overall, these algorithms manage to detect very well image modifications.