New Technique For Making JPEG Images Copy-Evident

Gunkerty Jeb writes "The days of wondering whether those drunken sex party photos are indeed the Olsen Twins, or if they are just the Mary-Kate and Ashley's faces photo-shopped on the bodies of Lindsay Lohan and Amy Winehouse are OVER! A group of academic researchers at the University of Cambridge has developed a new technique for making JPEG images copy-evident, so that users can tell whether an image has been recompressed and copied."

Comments

Image protected, but is it useful?

[Rashkae]

If the image can't be re-encoded or re-scaled without the watermark becoming visible, then it probably can't be resized for viewing either. So the only images can can really make use of this 'tech' are the ones that are already shrunk to their smallest desirable viewing size. I'm not sure how much use this will really have.

Re:Image protected, but is it useful?

[Hatta]

The only images you're going to want to protect are the ones you've distributed on the web.* Those are going to be the ones that are already scaled down. You do scale down your images before you host them, right?

*if you've decided to be a dick about copying, you're probably not going to be distributing your high res source image

Re:Image protected, but is it useful?

[Danh]

The original image can be resized without showing the watermark, see their demo page (and press Ctrl +/- if you browse with Firefox). But so can the re-encoded picture, it shows the watermark only at the zoom level of 100%.

From this I suppose that there is also one zoom level at which the original picture shows the VOID watermark (you better choose it to be an odd value)!

Re:Image protected, but is it useful?

[mobby_6kl]

I'm not sure what the method in TFA does since I didn't read it, but this isn't cold fusion.

Watermarking has been available for a while, even as a standard Photoshop filter. Just for a test, I applied the Digimarc watermark to a 18MP shot of my dog, and then resized it to 40% of the original size. The result? The watermark could be still read, albeit with a lower indicated watermark strength. Likewise, resaving it multiple times as a JPEG just slightly decreases the strength. If you aren't an idiot and don't start win a huge DSLR photo, it should be fine down to the common web sizes.

Of course if somebody knows about it and deliberately tries to destroy the watermark, I'm sure it won't be a huge problem. Like by applying a healthy dose of Gaussian blur.

Not ready for commercialization

[ColoradoAuthor]

From the original paper: "The technique now needs to be extended to handle arbitrary photographs, not just uniform regions."

Re:PrtSc

[gid]

Nice try at getting fast karma from a first post, but that's not going to work. The screen shot will already be of the compressed image, and will still show signs of re-compressing it.

Slightly Misleading

[sltd]

If you're really serious about putting Mary Kate and Ashley's head on Lindsay Lohan and Amy Winehouse's bodies, like it says in TFS, you'd use RAW. Then, you can compress everything together. Besides, if you upload that to photo sharing websites (especially Facebook) there's a high chance your picture would be recompressed, so it would have the compression artifacts whether it's been altered or not.

Fail.

Re:Slightly Misleading

[jeffmeden]

If you're really serious about putting Mary Kate and Ashley's head on Lindsay Lohan and Amy Winehouse's bodies, like it says in TFS, you'd use RAW. Then, you can compress everything together. Besides, if you upload that to photo sharing websites (especially Facebook) there's a high chance your picture would be recompressed, so it would have the compression artifacts whether it's been altered or not.

Fail.

Have access to RAW files of the aforementioned act involving Lindsay and Amy, along with ideally angled shots of the Olsen's faces? Didn't think so. So, you go to google images and start digging. What you find (in the ideal world this paper is picturing (ugh no pun intended)) is that every content creator (from the pr0n guys to the papparazzi that took the olsen's picture) have applied this filter to their work, and as such your efforts are for not. Don't get me wrong, there are tons of problems with this whole idea, but saying "well raw doesnt have artifacts anyway" is a bit of a fail.

Re:Slightly Misleading

[MobileTatsu-NJG]

If you're really serious about putting Mary Kate and Ashley's head on Lindsay Lohan and Amy Winehouse's bodies, like it says in TFS, you'd use RAW.

Uh, yeah, by 'use RAW' you mean: "take photos of these celebrities yourself".

...so it would have the compression artifacts whether it's been altered or not.

Unless you took the RAW photos yourself, there would still be different artifacts from each of the pictures you used to do the composite. More artifacts won't change that. Besides that, Facebook resizes the image THEN saves it, again altering the new artifacts in a not-so-subtle way.

You haven't shot this down at all and it's obvious that people with mod points don't understand this topic.

Re:Slightly Misleading

[amnesia_tc]

And "knot" is right out.

Re:PrtSc

[ElectricTurtle]

See, now I can tell it's a Photoshop from some of the pixels and having seen quite a few Photoshops in my time.

Doesn't Work

[nattt]

So I downloaded their test image here: http://www.cl.cam.ac.uk/~abl26/copy-google.jpg that they claim gets a message on it when compressed by google proxy http://www.google.com/gwt/x/i?u=http%3A%2F%2Fwww.cl.cam.ac.uk/~abl26/copy-google.jpg&wsi=223e8e5df695e99c&ei=6ixQTebOCoPoxQW8rYlv&wsc=yq&whp=012e012f72be

But when I take the original and re-save it in Photoshop CS5 I don't see the void lettering. I reduced the JPEG quality and kept trying and at quality 1, the lowest setting I was starting to see a pattern, but no words appeared.

I'd say their idea is nice, but doomed to failure, not least they mention "If you can’t see the message in the recompressed image, make sure your browser is rendering the images without scaling or filtering." which would be the obvious source of attack on such a method should it actually work in practise.

Re:Doesn't Work

[marcansoft]

Apparently it only works at very specific quality settings. Re-saving with GIMP, I can see the message at quality 24, 38, 41 (barely), 43, 60, 65 (barely), 69 (barely), 72 (barely), and a few others even less.

As far as I can tell (I haven't read the paper), it works by setting up a static hard to compress pattern, and then slightly altering that pattern in certain macroblocks so that they just push the boundary into a different kind of compression artifact at certain quality/quantizer levels. So the entire image is compressed one way at one quality, a different way at another quality, and at the threshold between them there's a quality level where the message blocks compress differently and you can see them.

Also, recompressing has a high chance of destroying the effect permanently. For example, saving at quality 51 destroys the message, and re-compressing at any quality level no longer makes it visible.

Re:Doesn't Work

[daremonai]

I tried it with gimp and it worked. Open the image, save at 60% quality, and the VOID is quite visible. Perhaps it's libjpeg's compression they're relying on. (I assume that's what gimp uses; I haven't actually checked.)

However, the method is easily defeated. Open the image, scale it to twice its size, then scale back down to the original size, and save at 60%. No watermark appears.

Re:social problem, technical solution

[madbavarian]

How long before browsers automatically low-pass filter these broken jpegs? If not, then the browser is going to have a hard time resizing the jpeg if this high frequency noise is so large in amplitude that it causes clipping in the jpeg calculations.

Re:social problem, technical solution

[Half-pint HAL]

Consider this: with several different HD sizes in frequent use, and even more computer screen geometries, for any given video you're going to find a lot of legitimate users viewing a radically resized version, so indiscriminate interference is unworkable. So who is this "specific combination" aimed at? YouTube. Could YouTube change their settings to get round this? Yes. Would YouTube change their settings to get round this? No. Because they would then be actively facilitating piracy and in breach of the DMCA. Any change of the YouTube reencoding settings would have to be thoroughly justified with some other goal, or they'd see themselves sued to bits.

HAL.

Re:not really new

[0100010001010011]

Error Level Analysis has been up for quite some time. It's not a hard set "this was a 'shop", but it's pretty good.

Re:PrtSc

[$RANDOMLUSER]

(keys desk intercom) Miss Jones, come in here! And bring your steganography pad!

Re:Scaling should not be affected

[b4dc0d3r]

If you read the article, they are depending on the JPEG compression artifacts for the watermark display. Resizing should not cause this because you are deling with the uncompressed image data.

Resizing and then saving as a JPEG will result in re-compression and the watermark appearing. Saving as anything else bypasses this completely.

This is only useful when you know what conditions will be applied. The example they give near the end, uploading to youtube, will apply only as long as youtube does not change their settings. Then you have to change your thresholds and all of your protected videos in the wold are unprotected.

Re:PrtSc

[mlts]

I want my secretary to use a one-time pad for transcriptions.

Re:PrtSc

[_0xd0ad]

Well, one of those palette entries can be transparent, so if you just stack enough GIF images on top of each other...

Okay... I keed...

Sort of...

NOTICE: This looks shopped!

[da3dAlus]

I can tell from some of the pixels and from seeing a few shops in my time. What, are they borrowing an algorithm from 4chan?

Re:PrtSc

[LocalH]

Isn't the whole point of the format that it knocks back the color pallet to ~256 colors?

Actually, no. GIFs can contain multiple image blocks, each with its own palette. Ever seen a GIF with just under 16-bit color? Check this shit out.

Re:social problem, technical solution

[coolsnowmen]

I don't think you understand what they mean by clipping in this sense. In JPEG compression, the DCT is performed, the matrix is formed, the coefficients are scaled, quantized(clipped), and then Losslessly compressed. The scaling/clipping(from quantization) IS the core of the low/med quality JPEG compression compression. It is not like "clipping" in the audio sense where it is horribly distorted due to over amplification, the quantization is intentional and in DCT space.

Also, as stated elsewhere on this forum, when a browser resizes a picture, the cache is not recompressed, which is when this intentional dithering becomes apparent.