How Watermarks Fail - CITP Blog
I wrote Wednesday about Randy Picker's suggestion of using digital watermarks to embed users' personal financial information into media files, to discourage users from sharing the files. Today, I want to talk more generally about watermarks and how they tend to fail. First, some background. Watermarks are subtle signals embedded in the background of media files. They are supposed to be unobtrusive but easy to detect if you know where to look. Different media have different kinds of watermarks. In a photo, the watermark might be hidden in subtle patterns of shading. In music, it might be in a very soft background buzz, or a barely audible echo. In many applications, a watermark must resist attempts by an adversary to remove it. For example, in Randy's scheme, a user might want to remove the identifying watermark from a media file because he wants to share the file illegally, or because he doesn't want his personal information exposed to cyber-intruders. It is often important to know how resistant a particular watermark is to removal. There has been plenty of research on this topic, from which we can draw lessons about how watermark removal tends to work. One theme is the power of Rosetta Stone attacks. The original Rosetta Stone was a stone tablet with the same text written in three ancient languages. This gave scholars who understood one of the languages a big boost in deciphering another one that they didn't understand. Similarly, watermarks tend to be defeated if an adversary can get his hands on a watermarked file, and the same file without the watermark. By comparing the two, the adversary can determine where the watermark lives, which is usually sufficient to remove the watermark from other files. Alex used this method in deciphering the MediaMax watermark (as described in our Sony CD DRM paper), and my colleagues and I used it also in analyzing the SDMI watermarks back in 2000. Almost as powerful as a Rosetta Stone attack is a comparison attack, where the adversary does not have an unwatermarked file, but does have the same file with several different watermarks in it. Any place where two of the files differ is a place where watermark information lives. Given several marked files, an attacker can locate all or most of the places the watermark is hidden, which is again the first step in removing the watermark. (In theory it might be possible to stop an adversary with access to a limited number of individually watermarked files from completely removing the watermark, if the watermark has lots of places to hide and is constructed cleverly. There is an interesting body of theory about how to do this and when it works. But in practice the assumptions underlying that theory rarely hold.) Even if the adversary cannot get access to multiple versions of a file (so that Roseta Stone or comparison attacks are not possible), he can usually still defeat a watermark if he has access to a device that can detect watermarks. By reverse engineering the device, he can figure out where it is looking for the watermark, which again puts him in a position to remove it. (Even if he can't dissect the device, he can use it as an oracle that tells him whether a particular file has a detectable watermark. Oracles are very helpful in attacking watermarks – Alex used one in his MediaMax watermark analysis, and my colleagues and I used one in our SDMI analysis.) All of this helps us to understand where watermarks are likely to be effective and where they're not. The best case for watermarking is where each file is published in a single version, with a watermark in a location that is not disclosed to the public and is not implemented in a device available to the public. This would hold true, for example, in a system that put a distinctive mark into all released versions of a file, and then looked for such watermarks in content broadcast on the radio or TV or downloaded from the net. Not nearly as strong is a system where there is a single watermark per file, and consumer devices check for the mark – it is subject to reverse engineering and oracle attacks. Weaker yet is a system where files are watermarked individually for each consumer – it is subject to comparison attacks. Weakest of all is a system where files are watermarked individually for each consumer and everyone is told how to read the watermarks. Here the adversary can use comparison attacks, and reverse engineering is not even necessary because the inner working of the watermark detector are well known. Alert readers will have noticed that all of the uses of watermarks for DRM (copy protection) seem to fall into the weak categories. That is because DRM applications require either that all devices check for the watermark – opening up reverse engineering and oracle attacks – or alternatively that a file be given separate watermarks for separate consumers – opening up comparison attacks. Watermarking has its uses, but it doesn't seem well suited for DRM.
