Rampart: Browser native on-device PII radaction | Hacker News Reader

Rampart: Browser native on-device PII radaction

ndstudio.gov

103 pointsnateb202238 comments

I have worked in this field and I am the author of this package - https://github.com/deepanwadhwa/zink

A few things jump out since this is done by the government:

- the lowest hanging fruit for this problem is to clearly tell people (citizens) not to share any personal info with chatbots which can cause financial harm or identity theft. the example on the page shows a person sharing their SNN with a chatbot to help them find an apartment - "My name is Maria Garcia, my Social Security number is 123-45-6789, and I make $1,950 a month. Can you help me find affordable housing?" - why?? this is the opposite of what i would expect a government to advise their citizens.

- it's never too late for a good policy; the government should have extended HIPPA and other data privacy laws to AI companies - the AI company must not store anyone's SSN, no matter how stupid the user is. It should be on the AI company to not store it; so this type of layer should be on the AI company's side.

- technical; there are quasi identifiers of privacy (that's what my package targets) that are asymptotically hard to to deal with - meaning - if you remove everything that can leak your privacy the text would become meaningless. i don't think rampart can solve for that either and it should be clearly said on the website.

I think all of your points are very valid, but it doesn’t remove the point that the government is trying to make it free and easy for application builders to do a little bit better than they are today. This is commendable on its own, even if it’s not perfect.

I find it interesting that the co-author of the OP violated long-standing security practices and copied a social security database into an insecure AWS instance without proper oversight.

Is it still commendable if in order to develop this, the taxpayer-funded developer themselves published our PII to a consumer, privately controlled third-party server without following federal data security protocols? Seems like an attempt to make up for the damage they caused.

Standing up your own AWS-equivalent internally is not much work, yet the PII data leaked to publicly available services due to this developer's personal choices and lack of oversight to prevent sensitive government data from being published externally.

Notably, this resulting work has not been published as public domain (CC0).

what i understood from the above comment is that the government should first do the basic policy making (set some non-negotiables for the ai companies etc) before jumping into solution building. let the market figure out the tech solutions.

Letting the market handle things is how you get all the problems America has today. Markets are a good way to distribute goods and services but they don't really work for public services with no money to be made. Things like roads, hospitals and open source software.

The purpose of a government is to serve its people, not make sure every business gets the same conditions at all costs. Picking winners and losers is fine if the losers still have a chance to beat the winners when they're really better.

If someone makes a free PII detector better than Rampart I bet the US government will either switch to it or make Rampart v2. For now it's enough. Perfect is the enemy of good.

Point is completely fair. It seems to be sort of a grey area that they skim over entirely on your point about sharing deep personal info.

Absolutely this. "PII" is a moving target that's heavily context-sensitive. Any entity that's large enough to care about PII almost certainly has some category of it that would be indiscernible from a handful of jumbled alphanumerics without a clear understanding of the context, both linguistic and organizational, in which the text sits. A general-purpose model would struggled to solve this in the broadest case, and will certainly be incapable of solving it in any specific case.

So that’s why big ballz (co author of the OP) put our PII in an insecure AWS instance via DOGE’s starlink terminal

https://www.csoonline.com/article/4046997/whistleblower-doge...

I have presented approaches like this to banking clients and they are still not very interested. The only thing that makes these people happy is zero data retention and deterministic redaction at the source. Regex over arbitrary string literals does not represent determinism in this context.

If your product is handling natural language conversations from end customers, there is not much you can do to prevent the occasional PII leak without ruining the rest of the pie. ZDR is your best mitigation if you actually want the magical AI experience to work the way the investors hope it can.

PII can often become disclosed by way of many correlated factors that are not considered PII on their own. Even a perfect AI system cannot capture all of these relationships. You could probably locate where I live within a 20 mile radius if you spent enough time analyzing my HN comments over the years. Not one of these comments on their own would trigger a PII filter.

> If your product is handling natural language conversations from end customers, there is not much you can do to prevent the occasional PII leak without ruining the rest of the pie.

For related reasons (e.g. rule compliance) I'm trying to convince folks to shrink the LLM's role to a kind of "guess which valid choice the user meant to ask for" level, and let the data-entry parts be small and deterministic pieces.

Alas, convincing anyone to use some symbolic GOFAI (or a strict "harness" that is 90% of the way there) seems hard these days. The allure of a promised magical silver bullet is strong.

I can't deal with "radaction".

98.4% is nowhere near good enough to call it PII redaction.

When a file is uploaded, we will make a best-effort attempt in your browser to remove any private details using Rampart[1]

[1] https://america.gov/privacy-policy

https://discuss.google.dev/t/trusted-automation-with-google-...

Plain text in a chat input is only one piece of the problem. What about files like PDFs and documents filled with PII.

Why did the National Design Studio see the need to put all the readable text on the right column of the page?

Because there's one part where you also get text on the left side so Claude thought it's okay to keep it as blank space for the rest. Or they wanted to have an active example on the left for the whole page and ran out of tokens.

I am Jack's complete lack of concern for the end user.

The skeptic in me really can't trust our current government to protect my information.

Good thing you don't need much trust in this case. Source available here - https://github.com/nationaldesignstudio/rampart

I suppose the model could be doing stuff, but you can also switch that out for your own with the source.

I mean, it can be run locally, so you don't necessarily have to trust it, unless there's been any model-as-a-vector CVE.

That hasn't happened yet, has it? Where running a model from HF directly compromises the machine as opposed to some breakout or exfil done by the model after the fact?

That said, if you're filing your taxes or have a 'real' ID, the government already has all pertinent information required to fuck you over either deliberately or via a leak, so...

The vulnerability to watch out for would be the model being trained to not redact some specific information.

You could never trust the US government to protect our information since it moved online.

Nothing to do with the fantastic megadrive game Rampart which is a shame.

What we'd really like is a PII redaction model for video...

what do you have in mind for video? isn't redacting faces/name plates quite mature already?

Is this good enough for HIPAA?

No.

It's the technique used by a lot of AI healthtech

What is? This is nowhere close to what HIPAA requires.

A lot of clinical decision support AI tools essentially scrub the PHI data on the frontend.

PHI is a different set of data points from PII amigo.

Same technique is good enough.

Which is why I was asking what you meant by "this technique."

If your question is: "Is client-side de-identification sufficient for HIPAA," then sure, assuming you are de-identifying the data elements required by HIPAA to be de-identified. You'd have to be a real big idiot to rely on something like this though.

If your question is: "Does this library, as described in the blog post, de-identify data sufficiently for HIPAA?" then no, it doesn't appear that it does.

If they're relying primarily on this type of tool to scrub PHI, I can tell you already they are not in compliance.

The source code should be public domain, no?

https://github.com/nationaldesignstudio/rampart ?

LICENSE errors out https://github.com/nationaldesignstudio/rampart/blob/master/...

CC by 4.0 is not the same as CC0 (public domain): https://creativecommons.org/public-domain/

Reply on news.ycombinator.com