Using AI to hide secret messages via Steganography

[thewayne]

Interesting stuff. I especially enjoyed the bits talking about older methods of hiding messages in plain sight, like marking words in print with invisible ink.

Steganography is an interesting art. It's not cryptography as the technically the text is plainly available - if you know how to read it. One method of steganography was encoding messages in photographs and then posting them online. There's lots of wasted bits in photos, so you alter the bits, which doesn't really alter the image, post the photo, the recipient knows how to decode the bits, the message is passed. But the technique is detectable because the image doesn't compress as well as an unaltered photo.

Detecting textual steganography requires that you analyze the message text and develop a word probability distribution. The word 'the' is one of the most commonly occurring words used in written and spoken communications, 'analysis' less so. By comparing normal text to steganographic text, you can make assumptions as to whether or not text contains a hidden message.

The text that the message is hidden IN is called the cover text. It might be something like a visit to a local museum, and then the AI will alter that text to inject your secret message. You can then send the altered message and the recipient can re-process it and extract your secret message.

Now, here's the interesting bit. By using AI, the difference in probability distributions can be reduced to zero. So an enemy - a censor, a hostile state actor, whatever - cannot accurately say that any given message contains stenographic text!

Word probability doesn't tell you what the hidden message is, just the likelihood of whether or not there is a hidden message there, which may mean an increased likelihood of a person or group coming under tighter scrutiny.

The problem that I see with this is they're talking about a "plug-in for an app like WhatsApp or Signal would do the heavy algorithmic lifting". I'm a little confused at this point. If they need to match the probability distribution of the cover text with the PD of the secret message, and it's done by an AI which is a supercomputer or a computer cluster, will you be able to do that with just a plugin on a smart phone? I'd like to see some more solid proof of concept here rather than 'our math models demonstrate' sort of stuff before human rights workers in bad places put themselves at risk with stuff like this.

https://www.quantamagazine.org/secret-messages-can-hide-in-ai-generated-media-20230518/

Comments

[disneydream06]

Dear James Bond,
We have a job for you. LOL!!!!!!!!!!!!!!!!
Hugs, Jon

Well ...

[ysabetwordsmith]

>> Steganography is an interesting art. It's not cryptography as the technically the text is plainly available - if you know how to read it. <<

If you want to be really secure then you stack your obfuscations. For example, first you put your message into a rare language. Then you put that into code. If you choose something like Morse code, it doesn't even look like a message unless someone knows exactly what to look for. Then you print that encoded message with invisible ink or steganography or whatnot. Then you insert that page somewhere secret, like inside the pasteboard cover of a book. And if you were a complete bastard about it, the original message was all kennings, archaic idioms, or references to memories that only you and your intended recipient shared in the first place.

Even if your enemies suspect there is a message somewhere, they will have one hell of a time locating it and then slogging through all the layers of protection, and honestly by the time they do that, it's probably outdated.

Re: Well ...

[thewayne]

One of the key concepts of cryptography is it doesn't necessarily have to be completely unbreakable, but hard enough to break that the message is now outdated.  If the message is "we attack in three days" and it takes ten days to break the crypto, then the crypto did its job. One of the most secure forms of crypto is one-time pads.  Sender and receiver each have matching OTPs to encrypt and decrypt messages, and each page is used once then destroyed.  Never reused.  Because the key is only used once, there's never a repeating pattern.  As long as the algorithm that generates the OTPs is horribly flawed and your adversary got ahold of it, it's pretty darn perfect.  Of course, if your adversary gets ahold of your messages and your used OTPs, you're probably screwed.

Re: Well ...

[ysabetwordsmith]

>>One of the key concepts of cryptography is it doesn't necessarily have to be completely unbreakable, but hard enough to break that the message is now outdated. If the message is "we attack in three days" and it takes ten days to break the crypto, then the crypto did its job.<<

I agree, that's an important point.

>> One of the most secure forms of crypto is one-time pads. Sender and receiver each have matching OTPs to encrypt and decrypt messages, and each page is used once then destroyed. Never reused. Because the key is only used once, there's never a repeating pattern. As long as the algorithm that generates the OTPs is horribly flawed and your adversary got ahold of it, it's pretty darn perfect. Of course, if your adversary gets ahold of your messages and your used OTPs, you're probably screwed.<<

You can create something like that using books. There's a code that relies on looking up letters or words in a book that both people have. If you use the same book, there's always a chance your enemy will figure out which it is, get a copy, and decode the message. But if you both have access to a shelf of books, which you move through using an agreed pattern, then the chance of an enemy cracking that approaches nil. It's hard enough if you're just using them from left to right. But you could just as well use them in the alphabetical order of Norse runes or Celtic ogham. If you're doing this in a section of library where nothing's been checked out in 30 years, and there are thousands of books, then you're right back to a code that's effective because it can't be cracked fast enough.

And if you're a complete bastard about it, you might even do this in such a way that using a particular wrong book to crack the code will yield a coherent but misleading message. Of course, that really takes a crypto genius.

[mtbc]

One challenge I've wondered about is embedded messages that have some defense against people trying to block them, e.g., by corrupting plenty of "low" bits in images or audio knowing that it won't perceptibly change how they seem. Of course, if you keep the messages short, it's statistically easier to still pull a damaged message out of the noise. though even then there are questions like if it's worth lengthening them by using an error-correcting code.

[thewayne]

I don't know how ECC works precisely, I understand the concept.  It would be easy to use a checksum of some sort to know if the message has been corrupted and you could  easily send a 'Resend message of X date/time'.  I would wonder if including ECC would potentially overwhelm the size of the embed and risk an image degradation. With VPNs and torrents and onion routers, there are ways to get messages through if you really want to work at it.

[silveradept]

It seems like LLMs make it easier to hide things in them, because they are already doing things according to machine patterns, and we tend to use machines to try and analyze large corpuses of material for their irregularities. Human codebreakers might be able to find it, but at that point, it's done the job of being too expensive to crack.