I'm just speculating here, and I don't know how well it would work in practice, but you could do it if you limit to specific sources of tampering. E.g., suppose you want to remove the logo added to an image by memegenerator.net.
You know in advance what the text looks like and where it is. Create a transparent png template that matches the text. Then sum the differences between the image and template pixel colors, multiplying each by the alpha of the template pixel. Since for this particular logo, it's basically white (although it seems to have a thin black shadow) you would get false positives for a picture with a white part there, so you'd also need to verify that the surrounding pixels are (within a tolerance) not white. It's not clever but it could work for certain sites.
For anything more flexible (e.g., logos on images which have subsequently been resized) you're in to the territory of OCR and TinEye-like image matching, which are more advanced than I could advise you on.
To correctly detect all kinds of "tampering" and filter "illegitimate" from "legitimate" in general, you'd need an artificial intelligence that could understand the meaning and context of what it's seeing. The short answer is: you can't. That's what humans are for.
If this is for a website, probably the best thing you can do is a report button that lets users of your site report images that don't fit with your site's rules.