Many projects ask to share lots of logs when reporting issues. It’s difficult to go through all the logs and redact informarion such as usernames, environment variabled etc.
Any ideas on how to anonymize logs before sharing? Change your username to something generic?
The problem is there’s likely not a universal solution that’s guaranteed to clean everything in every case.
Cleaning specific logs/configs is much easier when you know what you’re dealing with.
Something like anonymizing a Cisco router config is easy enough because it folllows a known format that you can parse and clean.
Building a tool to anonymize some random logs from a specific software is one thing, anonymizing all logs from any software is unlikely.
Either way, it should always be double-checked and tailored to what’s being logged.
I wonder if you could do something with heuristics or a micro LLM to flag words that might be expected to be private.
I would be curious if someone could do a proof of concept with the Ollama self-hosted model. Like if you feed it with examples of names, IP addresses, API-key-like-strings, and others, it might be able to read through the whole file and then flag anything with a risk level greater than some threshold.
I agree, besides basic patterns to search for, that will most likely be necessary. In fact looking a bit more at this tool, it has a list of “rules” tailored to each software specifically, I guess this could be sustainable really only if a repository of third party extensions was kept so that anyone could contribute and the pool of rules expanded progressively