I think it is possible to embed invisible information into videos and images. This way peopple could track where you got an image from, the source from which you copied it, and people who copy your image to share it again. https://github.com/ShieldMnt/invisible-watermark
Services like youtube or twitter could embed such watermarks into content they serve to specific users without them knowing; Smartphone-cameras could mark images in secret.
I guess blurring, rotating or dithering the image could destroy watermarks. Or maybe just sharing a screenshot of an image instead of the original image. Format conversions may help too.
Keywords: digital-watermarking. tracking.
It sure is possible to embed invisible information into videos and images, it’s called metadata. Now you might think of other techniques, e.g. https://en.wikipedia.org/wiki/Steganography but most if not all are, AFAIK (and I won’t pretend I know the state of the art in the domain) if they are within the data itself (thus become data, not meta-data), e.g. a visible stamp in an image, are made to remain visible. Compression codecs are specifically targeting the visible or audible spectrum. One of the most basic way to “compress” lossy information (as opposed to lossless) is precisely to remove the ends of the spectrum that is not perceived by the average human audience.
So… AFAICT it’s either visible and thus can be spotted (and thus can be removed, even if by adding a black mark over) or not visible but then most likely will be removed by basic compression codecs even without trying to do so.
TL;DR: no and I wouldn’t be until I see this in the wild (not a research paper claimed it’s technically possible).
Pretty concerned. I have access to copyrighted content that I want to share elsewhere, but because it requires a log in, I always wonder if there are invisible trackers hidden in the files that would out me. I think it should be standard that all sites strip things like exif data automatically upon upload. Hexbear does this.
I didn’t think of this. Can they embed dynamic info that logs ip and identifying info?
Yes but encoding isn’t cheap.
If I’m sharing images, I’m generally sharing links to images, or images I’ve created myself and scrubbed of metadata, plus touched up, usually with a noise mask, among other things.
Most people don’t know your photos can be cross profiled and identified by the unique noise signature of your camera.
I’ve never heard of it being used in practice though. There’s a github repo somewhere if you’re interested in trying it yourself.
maybe blur + compression + dithering + contrast effects may fix this? idk …
I would be interested in the name of the github repository you mentioned.
I believe this is the original whitepaper: https://ws.binghamton.edu/fridrich/Research/double.pdf
And here’s an implemention I found on github: https://github.com/andrewlewis/camera-id
With that repo you should be able to test ways to obfuscate the noise signal.
Then that could be used to fingerprint too.
It’s something to be aware of if you are trying to record/send images in secret. But it’s not a problem if you are passing on an image you found somewhere else. You leave no trace if you are just passing the image, so the image itself is of no concern privacy-wise.
Wouldn’t recoding or using other compression methods break that?
Not necessarily, a qr-code can still work even if you compress it, due to digital error-correction mechanisms in the code.
That is built for purposes vs trying to hide something.
One might think that watermarks are designed to resist corruption as well. But I have no idea.
They are, if you scroll to the bottom of the github repo that OP posted there are some examples of what works and doesn’t work to break it.
Watermark data like this is stored in the least significant bits of the pixels themselves, or in the case of OPs example, they do a frequency decomposition on the image then store the watermark data in the coefficients. Basically you have to trash the pixel data at least a little bit to defeat it. So cropping or flipping the image won’t do it, but resizing or rotating with some kind of filtering will.
I have no idea how the machine-learning technique listed there is working, and their documentation link is broken :(
