oxeo0/opsec-blogposts
1
0
Fork 0
mirror of http://git.nowherejezfoltodf4jiyl6r56jnzintap5vyjlia7fkirfsnfizflqd.onion/nihilist/opsec-blogposts.git synced 2025-06-08 16:09:33 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

add rewritten stylometry tutorial

Browse source
This commit is contained in:
oxeo0 2025-06-07 11:57:07 +02:00
parent 0838f429c7
commit 0154812c4c
No known key found for this signature in database
GPG key ID: B4DCEAB52B5BEC67
6 changed files with 5 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

8
stylometry/index.md
View file

@ -1,6 +1,6 @@
---
author: zl
date: 2025-05-31
date: 2025-06-07
gitea_url: "http://git.nowherejezfoltodf4jiyl6r56jnzintap5vyjlia7fkirfsnfizflqd.onion/nihilist/blog-contributions/issues/13"
xmr: 83geT3KQZGthZ99r1z72t58TFztdDHGHjgnCB3jvniV8FC1bcYf6HriDnSpnt2SZXzcBByNCcjRHiPmtNu5G8CuNG9mfDyY
tags:
@ -46,7 +46,7 @@ This identifiability provides more direct attack vectors for a resourceful adver
This approach has been used in [many areas](https://en.wikipedia.org/wiki/Stylometry#Case_studies_of_interest) like forensics, authorship attribution, or determining if [Shakespeare](https://www.latimes.com/science/sciencenow/la-sci-sn-shakespeare-play-linguistic-analysis-20150410-story.html) wrote a play.
Multiple anonymous personas, even if they are fully [segmented](../internetsegmentation), can also potentially be correlated based on writing styles, leading to information from both being combined for deanonymization.
Multiple anonymous personas, even if they are fully [segmented](../internetsegmentation/index.md), can also potentially be correlated based on writing styles, leading to information from both being combined for deanonymization.
Importantly note that stylometry does not have to be done to this scale to be usable. Even without profiles, an adversary can extract a lot of information with algorithmic analysis. As discussed, each minor detail of someone's writing will always give insight into who they are.
@ -70,7 +70,7 @@ By using LLMs, the writing details will be inaccurate to the original text and u
### Software and Requirements
For this tutorial, we will be using **Ollama** for the backend and **Open WebUI** as the frontend. Instruction on how to set both of these up along with information on model choice and performance are in the [local llms tutorial](../openwebuilocalllms/), which is a recommended prerequisite. This post can still be followed with alternate backends or frontends.
For this tutorial, we will be using **Ollama** for the backend and **Open WebUI** as the frontend. Instruction on how to set both of these up along with information on model choice and performance are in the [local llms tutorial](../openwebuilocalllms/index.md), which is a recommended prerequisite. This post can still be followed with alternate backends or frontends.
Here is a visual of what that setup looks like:
@ -178,7 +178,7 @@ Using the same process as before, we can save this as a preset and call it Gemma
![image of friendly rephrasing](15.png)
It works, the model outputting lowercase text with friendly, outgoing wording, making the personality of the message completely different and obfuscating the original tone entirely.
It works, the model outputting lowercase text with friendly, outgoing wording, making the personality of the message completely different and obfuscating the original tone entirely.
Now let's try another writing style, going for short rephrasing. We will write the prompt slightly differently. Creating a preset, we can call it Gemma-4B:Short: