Predicting the text in redacted documents is close to reality
Releasing delicate information with big black bars all over it has kept secrets safe for years - but not for much longer, maybe.
For those with secrets they want to keep, redacting documents is a pretty important thing to get right. It’s necessary to understand how to redact documents, firstly - look to Southwark Council, which in February uploaded its controversial agreement with developer Lend Lease for the regeneration of the Heygate Estate in a form that let people copy and paste the text underneath the black bars.
But it’s also necessary to know which parts of a document to redact so that the context from the stuff left open doesn’t give the game away. There is always, however, information left behind. The choices made in how to block text - be it with other bits of paper, or black marker pen, or even by typing out new words and then covering those up - can reveal something about the person doing the redacting. Different agencies had different redaction standards at different times, which gives a further clue as to what technique is needed. Each typeface fits into the space under a bar in a limited number of contextually-relevant ways, as well.
In the New Yorker, William Brennan reports on The Declassification Engine, an intriguing attempt by a group of academics to use these clues to try and crack any redacted text. A snippet:
Together with a group of historians, computer scientists, and statisticians, [Columbia history professor Matthew] Connelly is developing an ambitious project called the Declassification Engine, which, among other things, employs machine-learning and natural language processing to study the semantic patterns in declassified text. The project’s goals range from compiling the largest digitized archive of declassified documents in the world to plotting the declassified geographical metadata of over a million State Department cables on an interactive global map, which the researchers hope will afford them new insight into the workings of government secrecy. Though the Declassification Engine is in its early stages, Connelly told me that the project has “gotten to the point where we can see it might be possible to predict content of redacted text. But we haven’t yet made a decision as to whether we want to do that or not.”
One of the things that jumps out in here is the parallel between the "mosaic theory" - where "pieces of banal, declassified information, when pieced together, might provide a knowledgeable reader with enough emergent detail to uncover the information that remains classified" - and critics of the NSA who realise that mass collection of metadata rather than the actual data of communications is, in many ways, just as bad.