Introducing the NSA-Proof Crypto-Font

Daniel_Stuckey writes "At a moment when governments and corporations alike are hellbent on snooping through your personal digital messages, it'd sure be nice if there was a font their dragnets couldn't decipher. So Sang Mun built one. Sang, a recent graduate from the Rhode Island Schoold of Design, has unleashed ZXX — a 'disruptive typeface' that he says is much more difficult to the NSA and friends to decrypt. He's made it free to download on his website, too. 'The project started with a genuine question: How can we conceal our fundamental thoughts from artificial intelligences and those who deploy them?' he writes. 'I decided to create a typeface that would be unreadable by text scanning software (whether used by a government agency or a lone hacker) — misdirecting information or sometimes not giving any at all. It can be applied to huge amounts of data, or to personal correspondence.' He named it after the Library of Congress's labeling code ZXX, which archivists employ when they find a book that contains 'no linguistic content.'"

Comments

Re:Easy to crack?

[dgatwood]

Depends on the steganography method used, and on how many images are sent using that method. If you're a spook and you see somebody suddenly sending lots of images to someone else, you might grow suspicious, at which point you'll start performing analysis to see if there are patterns emerging across the entire set of images, such as certain pixels that are always higher than the adjacent pixels by a certain amount. Granted, such patterns can just as easily be caused by sensor flaws, but some fairly primitive steganography techniques could be detectable in this way.

Second, because subpixel noise in cameras isn't random—it tends to obey a gaussian distribution, and thermal noise can vary considerably from frame to frame depending on the length of the exposure—when spread over a large enough number of sequential or nearly sequential photos taken by the same camera, the steganography might be detectable by using a model of the predicted levels of noise that the image sensor should produce for a shot of a given duration and the elapsed time since the previous shot. This won't tell you what is embedded in the image, but if you're lucky, it might tell you that with a high probability, something is embedded. Depending on the circumstances, that might be enough to get a warrant. Then again, it could just be Digimarc.

Finally, there's the question of the randomness of the source material (or, more to the point, the lack thereof). If the base image is at the native sensor resolution of the camera, the nature of the image sensors themselves could potentially be exploited to detect some types of steganography. In a real-world image sensor (except for Foveon sensors), there's no such thing as a pixel; there are only subpixels that produce a value for a single color. The camera must combine these values (a process called "demosaicing") to compute the color for a pixel in the final image. Because the subpixels that make up a pixel are not physically on top of one another, the camera typically computes the estimated value for the color at a given physical point on the sensor by combining adjacent subpixel values in differing percentages. For example, if the green subpixel is chosen as the "center" of the pixel and the red subpixel is to the left and the blue is above, it might mix a bit of the red from the "pixel" to its right and a bit of the blue from the "pixel" below it. (This explanation is overly simplistic, but you get the basic idea.)

Unfortunately for steganographers, the way that particular cameras construct a pixel value from adjacent subpixel values is predictable and well understood. If a steganographic technique does not take that into consideration, it is highly likely that, given knowledge of the camera and its particular mixing algorithm, the steganographic data can be detected simply by determining whether there is any plausible set of subpixel values that could result in the final computed pixel values for the entire image. For that matter, given that most of the algorithms for subpixel blending are straightforward, even without knowledge of the particular camera, it is highly likely that steganography can be detected, because portions of the image that contain no hidden data will likely only be producible by a single algorithm, and portions of the image that contain hidden data likely will not be.

Those are just a couple of types of analysis off the top of my head that might potentially be used against some types of steganography, given some types of source material, etc. It is entirely possible that there are steganographic techniques that are resistant to these sorts of analysis, and there are likely many other interesting types of analysis that I have not mentioned. I have not kept up with steganographic research personally, so I can't say with any certainty.