How to defeat naive image steganography | Hacker News Reader

How to defeat naive image steganography

incoherency.co.uk

47 pointsjstanley29 comments

There is a less naive steganography tool in the Debian repos named steghide. According to the manual:

       Steghide uses a graph-theoretic approach to steganography. You do not
       need to know anything about graph theory to use steghide and you  can
       safely  skip  the rest of this paragraph if you are not interested in
       the technical details. The embedding algorithm roughly works as  fol‐
       lows:  At  first, the secret data is compressed and encrypted. Then a
       sequence of postions of pixels in the cover file is created based  on
       a pseudo-random number generator initialized with the passphrase (the
       secret data will be embedded in the pixels at  these  positions).  Of
       these  positions  those  that do not need to be changed (because they
       already contain the correct value by chance) are sorted out.  Then  a
       graph-theoretic matching algorithm finds pairs of positions such that
       exchanging their values has the effect of embedding the corresponding
       part  of  the secret data. If the algorithm cannot find any more such
       pairs all exchanges  are  actually  performed.   The  pixels  at  the
       remaining  positions (the positions that are not part of such a pair)
       are also modified to contain the embedded data (but this is  done  by
       overwriting  them,  not  by  exchanging them with other pixels).  The
       fact that (most of) the embedding is done by exchanging pixel  values
       implies  that  the first-order statistics (i.e. the number of times a
       color occurs in the picture) is not  changed.  For  audio  files  the
       algorithm  is the same, except that audio samples are used instead of
       pixels.

I wonder how hard it would be to detect (not decode) content hidden by steghide.

There is a series of lectures on stego at Oxford [1]. The lecturer uses simple machine learning (average perceptrons I think) and achieves 99% accuracy with little effort although that was for LSBR[2]. Later he discusses problems with the graph approach [3]. Simply put whilst it preserves 1st order histograms it will probably mess with 2nd or 3rd order histograms.

[1]https://www.cs.ox.ac.uk/teaching/courses/advsec/ [2]https://www.cs.ox.ac.uk/teaching/materials15-16/advsec/advse... p48 [3] ^ p58

Your second link is behind authentication. Is it freely available somewhere?

I always expected steganography to be used to encode a compressed and encrypted datastream, not uncompressed images. Such a compressed and encrypted datastream should be indistinguishable from noise, so not show in this kind of analysis.

My pet idea for doing steganography was to use nonstandard or suboptimal encoding sequences. For compressed data formats, including images, music, video, and general files, there is almost always a choice of various encodings which will decompress to the same result. Systematically varying a compression choice gives an invisible way to encode data.

Actually there are properties of images that meant that the noise introduced through stego is actually, especially for LSBR, VERY easy to detect. https://www.cs.ox.ac.uk/teaching/materials15-16/advsec/advse...

That page requires an Oxford University username and password.

I'd be interested to read this.

You too? That's #2 on my retirement list after releasing a library of open source epistemology enhancements for common game engines with demo UI. Maybe, together, we can finally kill the dialogue tree.

I wrote something like that for an ARG. Basically I used GIF images with sets of 2 visually identical entries in pallet and encoded bits by picking either the first or second. Not quite as sophisticated as something like varying which dct coefficients to keep in a jpeg, but it was intended to be breakable with a hex editor.

The problem was that most image hosts were smart enough to detect the sub-optimal encoding and "helpfully" fix it.

About 2 years ago I have implemented the Steganalysis method based on paper "Steganalysis by Subtractive Pixel Adjacency Matrix" http://www.ws.binghamton.edu/fridrich/Research/paper_6_dc.pd.... The whole evaluation was done by embedding random messages and the method worked well so I guess encryption of the datastream won't help much.

Wouldn't this be completely defeated by just simply encrypting the data beforehand?

Yes.

And, in general, people don't hide the image of some text inside another image. They hide the text. And that isn't so linear (blocks of 0s or 1s)

But we agree that the bar for something non-naive is extremely high if you want to defeat the well-funded adversaries and whatever they come up with in the next few decades (future algorithms will be applied to today's images). I'm sure they have nice libraries of the artifacts most cameras and compression algorithms should leave in any image and if it isn't there, well, let's look for some steganography?

I'm pretty sure you are overestimating the capabilities of government agencies. At least I hope so.

Ha! In general I assume that the NSA is ahead in many areas that remain classified.

I liked the idea... yes, it's possible to have a catalog of expected artifacts in JPEG images, in general. And the other way it would also make a good place to hide: changing artifacts from one camera to another, for example.

But the text would still be super-easy to find because it would create a bias that's not normally present in the least significant bits. Also, if not encrypted, the bias would conform to the letter frequency of one of the natural languages.

Think about captcha and why you might use an image of a cyphertext to defeat automated tools.

Could be combined with barcodes as well.

The title specifies that it is talking about naive steganography. So, yes, your less-naive implementation would thwart this analysis :)

Yeah, normally it would be encrypted text. Text is not too large. Encrypted so "random".

This is why the steganography tool PixelJihad uses a schedule to scatter pixels randomly around the images, instead of changing LSB sequentially (and derive the same schedule to retrieve the pixels: https://sekao.net/pixeljihad/about.html )

Definitely not naive, but an interesting hack.

I have been thinking a lot about stenography. If we can make the technology a little more common place ( couldn't find a good book on the topic ) - it would effectively make the arguments made by FBI,GCHQ ineffective. But something tells me large scale usage of stenography is not simple to implement.

There are two big problem with stenography: 1) as it is inefficient, you need a "normal"/casual way to send a big amount of data. It used to be rare, but nowadays I doubt that this is a big issue. 2) you need to coordinate how you're going to use stenography. That is the real big issue, same as key distribution in cryptography.

When it was reported that OBL had a big porn stash in the house at Abbottabad I presumed that it was (at least partly) to be used with steganography.

Nice.

Here's some messages hidden in audio 'noise' generated from stretching out all the black and white squares of a QRcode and generating some audio from it.

It can also be decoded back to the QR code and back to the message.

Neat.

http://qrdio.com

Just realised I obviously cocked up the bit planes for the Lena+Walrus example. There's some grey in there as well as black and white!

EDIT: Fixed. It was because I changed it down to greyscale too late in the process.

Knowing nothing about steganography, how far would I get from this detection by making the last bit-plane the xor or the second last one and my data?

I've not tried it, but it seems like it could work. The only potential problem is that you'll have around the amount of noise expected in the (n-1)th bit instead of the nth bit.

I've posted it elsewhere in the comments but Dr. Andrew Ker at Oxford Uni runs a course on steganography [1]. It's worth a read as he covers several different methods for stego and then methods for detecting said methods.

[1]https://www.cs.ox.ac.uk/teaching/materials15-16/advsec/advse...

"This service is accessed via the University of Oxford Single Sign-On system."

Are you able to exfiltrate this information?

Reply on news.ycombinator.com