Suramya's Blog : Welcome to my crazy life…

August 31, 2023

Using LLM’s to change writing style to hide author?

Filed under: Emerging Tech,My Thoughts,Tech Related — Suramya @ 12:17 PM

It is fairly well known that folks can identify a writer based on their writing style. In fact there is a whole field of work called Stylometry that analyses writings to try to attribute authorship of documents/writings. This is used in when new text/writings are found that are not attributed to any person or when agencies receive letters or there are articles posted on extremist sites.

In 1964, Frederick Mosteller and David Wallace published a three-year study of the distribution of common words in the Federalist Papers and showed that the writing style of Alexander Hamilton and James Madison differed in subtle ways. The researchers analyzed some disputed papers that both writers claimed to have written and based on their analysis of the respective writing styles of both authors were able to identify that the papers were likely to have come from Madison’s pen. Since then the technique has only improved, with AI and Machine learning algorithms coming into play researchers are able to identify the writing style from a post quickly and mostly accurately. Keep in mind that this is still more of an art than a science. However, it works well enough that people were able to identify the author of a book: “The Cuckoo’s Calling, by Robert Galbraith” as none other than JK Rowling.

Now, as we all know “With every action there is an equal and opposite reaction” and there has been a lot of work put into what is called Adversarial stylometry. Which is the practice of consciously changing the writing style when writing so remove the identifying characteristics that can be used to figure out who wrote the text. Historically, this was achieved by paraphrasing the text and by having another person edit & modify the content.

With the recent advances in LLM especially systems like ChatGPT it might be possible to use them to whitewash the text and remove enough identifying marks that the text becomes relatively more anonymous. I tried it out with the previous paragraph as the input and got the following as the result:

As commonly understood, the principle of “For every action, there is an opposing reaction” holds true. Considerable effort has been dedicated to a field known as Adversarial Stylometry. This involves intentionally altering one’s writing style to eliminate distinctive traits that could be employed to determine the authorship of a text. In the past, this was accomplished by rephrasing the content and enlisting the assistance of a third party to revise and adjust the material.

As you can see the output is significantly different in both words, tone and phrasing that it would it hard to attribute correctly. Do it a few times in a loop and after ensuring the original meaning is not lost you can use it and be relatively sure that it will be difficult to identify the author of the text.

Obviously, if LLM’s are going to be used to obfuscate text then other LLM’s will be created that try to de-obfuscate the text and the systems will keep getting more and more smarter.

– Suramya

No Comments »

No comments yet.

RSS feed for comments on this post. TrackBack URL

Leave a comment

Powered by WordPress