Top Secret: Automatically filter sensitive information(thoughtbot.com)
thoughtbot.com
Top Secret: Automatically filter sensitive information
https://thoughtbot.com/blog/top-secret
15 comments
> - NER doesn't often perform well in different domains, how accurate is the model?
https://github.com/mit-nlp/MITIE/wiki/Evaluation
The page was last updated nearly 10 years ago.
https://github.com/mit-nlp/MITIE/wiki/Evaluation
The page was last updated nearly 10 years ago.
When I had to implement "deidentification" for a kind of sensitive safety reporting, an LLM would've been a good way to augment the approaches I used.
Today, if I had to do it, I'd probably throw multiple computer approaches at it, including LLM-based one, and take the union of those as the computer result, and check it against a human result. (If computer and human agree, that's a good sign; if they disagree, see why before the document goes where it needs to be deidentified.)
(In some kinds of flight safety reporting, any kind of personnel can submit a report about any observation related to safety. It gets very seriously handled and analyzed. There are also multiple ways in which the reporting parties are protected. There are situations in which some artifacts need to have identifying information redacted.)
Today, if I had to do it, I'd probably throw multiple computer approaches at it, including LLM-based one, and take the union of those as the computer result, and check it against a human result. (If computer and human agree, that's a good sign; if they disagree, see why before the document goes where it needs to be deidentified.)
(In some kinds of flight safety reporting, any kind of personnel can submit a report about any observation related to safety. It gets very seriously handled and analyzed. There are also multiple ways in which the reporting parties are protected. There are situations in which some artifacts need to have identifying information redacted.)
Oh hey! Good to see this. I built something similar in python a while ago.
Check it out: https://github.com/deepanwadhwa/zink
The shield functionality fits directly in your LLM workflow.
Check it out: https://github.com/deepanwadhwa/zink
The shield functionality fits directly in your LLM workflow.
There is an extension for PostGres, https://postgresql-anonymizer.readthedocs.io that allows you to mask data by user or group at the schema level with the options to return full mask, partial mask or dummy data.
Can filters of this type be used when doing a screenshare or streaming? E.g any application that is on the screen.
Would be good to not have to worry about leaking a lot of stuff to the world.
Would be good to not have to worry about leaking a lot of stuff to the world.
I think there is a difficult cost benefit analysis with screen sharing. For example, what constitutes PII in that context? Maybe the text on the screen can be extracted and redacted but what about peoples faces? What about website icons in your browser tabs lists? It feels like eventually you’d need to redact everything to have privacy or accept loads of leakage
If you can run inference in real time (while doing a video call), and you can extract text through your operating system's accessibility APIs (eg the application isn't doing it's own bespoke text rendering), then probably. You'll still need to figure out where the entity appears on screen in order to censor them. (Or maybe the accessibility APIs have that information?) And you'll need some way to get in-between the OS and the screen share, like a virtual display or something.
How can one use this for logging? Won't the LLM performance lag grind your application to a halt?
They're using MITIE[1] models for NER. And MITIE models aren't LLMs. They're way lighter: word embeddings with SVM classifiers.
[1]: https://github.com/mit-nlp/MITIE
[1]: https://github.com/mit-nlp/MITIE
US Marshalls told me they need tech like this, automatic redaction.
I assume the NER model is small enough to run on CPU at less than 1s~ per pass at the trade off of storage per instance (1s is fast enough in dev, in prod with long convos - that's a lot of inference time), generally a neat idea though.
Couple questions;
- NER doesn't often perform well in different domains, how accurate is the model?
- How do you actually allocate compute/storage for inferring on the NER model?
- Are you batching these `filter` calls or is it just sequential 1 by 1 calls