Comment by mlissner
14 hours ago
Cool to see this here. It’s funny because we do so many huge, complex, multiyear projects at Free Law Project, but this is the most viral any of our work has ever gone!
Anyway, I made X-ray to analyze the millions of documents we have in CourtListener so that we can try to educate people about the issue.
The analysis was fun. We used S3 batch jobs to analyze millions of documents in a matter of minutes, but we haven’t done the hard part of looking at the results and reporting them out. One day.
https://www.argeliuslabs.com/deep-research-on-pdf-redaction-...
> Information Leaking from Redaction Marks: Even when content is properly removed, the redaction marks themselves can leak some information if not done carefully. For example, if you have a black box exactly covering a word, the length of that black box gives a clue to the word’s length (and potentially its identity).
Does X-ray employ glyph spacing attacks and try to exploit font metric leaks?
No, we worked with researchers that developed that kind of system, but didn't broadcast our work b/c the research was too sensitive. Seems the cat is out the bag now though.
I think the combination of AI and font-metrics is going to be wild though. You ought to be able to make a system that can figure out likely words based on the unredacted ones and the redaction's size. I haven't seen any redaction system yet that protects against this.
I thought glyph spacing attacks are an old idea; like I recall reading about such ideas 10-20 years ago unless I’m misremembering. Can you clarify why it was considered “too sensitive” if the whole point of this effort is to showcase these attacks?
This is going to be a disaster IMO because AI will just hallucinate what it thinks is the most probable redacted word and people will take that as gospel.
> I haven't seen any redaction system yet that protects against this.
The linked article suggests widening redacted areas more than needed with some randomization applied to the width. Strikes me that that wouldn't do much except add a few more possible solutions.
4 replies →
Presumably with font kerning and pixel perfect recreation of the source, it would be possible to guess the word very accurately.
The strings oioioi and oooiii will have different widths in some fonts because character organisation matters a lot.
I suppose it gets a bit more complex again if you enable stuff like microtype, but even then you can probably measure how much inter-letter and inter-word spacing has been adjusted by just scanning other text in the same line.
I think the conclusion is honestly that PDF is an outdated format for keeping records that might have to be redacted in the future, like court documents. Something reflowable like epub could have the text replaced with constant-space black squares instead no hints leaked as someone mentioned in a parallel comment.