Comment by mlissner
16 hours ago
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?
It’s a fine line. Most redactions are for the good, to protect someone or something. For example even in the Epstein files, where some redactions are being abused, most redactions are protecting victims.
If there’s a way to undo huge amounts of redactions, that’d certainly be a net negative. Sort of like if encryption were suddenly broken, you wouldn’t publish a paper saying so.
Our goal has always been to educate about the problem so that it can be addressed. We didn’t have resources to push on the font metrics approach, so we stayed mostly quiet about it.
> If there’s a way to undo huge amounts of redactions, that’d certainly be a net negative. Sort of like if encryption were suddenly broken, you wouldn’t publish a paper saying so.
I can't state emphatically enough how this is not the right mental playbook.
If you have found a vulnerability, it's likely someone else has too. By sitting on it, you only create more future victims.
Disclosure will lead to fixing this issue, mitigating it's precense, or switching tools/workflows, possibly a combination of. Sitting on it only ensures that folks who think they are protected, actually aren't.
> 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.
Yeah, the more robust protection is to widen to a constant. But in the general case that could require reflowing the pdf. But honestly single word redactions are really probably useless with cheap AI that can highly accurately fill in the gaps
Depends what you're trying to hide.
If the redaction is a person's name, and there's nothing else to give the person's identity away, single word redaction probably works reasonably well, AI or no AI.
3 replies →
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.
"don't redact or we will hallucinate something worse and make people believe it as gospel" is nice deterrent
We don’t need a “deterrent” against things being redacted in publicly released documents. We can have transparency without the whole world finding out the names of victims and witnesses, people’s phone numbers and SSNs, etc., every time a document is released.