This repository contains the code and the scripts to reproduce the experiments from the paper StyleRemix: Interpretable Authorship Obfuscation via Distillation and Perturbation of Style Elements.
StyleRemix, is an adaptive and interpretable obfuscation method that perturbs specific, fine-grained style elements of the original input text. StyleRemix uses pre-trained Low Rank Adaptation (LoRA) modules to rewrite inputs along various stylistic axes (e.g., formality, length) while maintaining low computational costs.
We use StyleRemix for obfuscation on four datasets: presidential speeches, fiction writing, academic articles, and diary-style writings. StyleRemix outperforms state-of-the-art authorship baselines and much larger LLMs on an array of domains on both automatic and human evaluation.
In this repo, we provide code which implements StyleRemix on a Llama-3 8B model for these four datasets.
To set up the environment to run the code, make sure to have conda installed, then run
conda create --name obf python=3.10
Then, activate the environment
conda activate obf
Finally, install the required packages (make sure you are in the root directory).
pip install -r requirements.txt
We include the python file quickstart.py which shows how to run StyleRemix on some sample texts. This uses LLama-3 8B as the base model, so you must do inference with a GPU. Style element weights can either be randomly set or randomly initialized; see the argparser for the full options. Finally, feel free to replace the sample texts with your own data.
You can either import the remix(...) method into your own code, or you can run this file directly, for example:
# Example commands to run StyleRemix on text:
# Passing in manually weights for different style elements (higher=more, lower=less, -1 to 1)
python3 quickstart.py --length 0.7 --sarcasm 0.9
# Randomly set weights for different style elements
python3 quickstart.py --random_weights --num_random 3
All resources (trained models, demo, etc.) have been organized into the following huggingface collection.
We use the AuthorMix data which is composed offour different domains, presidential speeches (curated in this paper), fictional novels (curated in this paper), the Extended-Brennan-Greenstadt (Brennan et al., 2012) (amt) and the Blog Authorship corpus (Schler et al., 2006) (blog), using a range of different authors (3 - 5). All raw datasets can be found under the test_data/ folder. Note the file test_data/AuthorMix is a torch file with a dictionary containing a key for each domain (Speeches, Novels, AMT, Blog) and the file test_data/AuthorMix_average_by_author contains a pre-computed matrix of average automatic evalution by author which is used to choose the weights of the adapters.
The test dataset can also be downloaded directly from huggingface: link
from datasets import load_dataset
data = load_dataset("hallisky/AuthorMix")
Experimental code for both all domains can be found in the main folder labeled as obfuscate.py. Each experiment consists of the following four steps:
- Download Raw Data: automatic
- Run Automatic Evaluation: automatic if using current domains
- Choosing Styles to Perturb for Adapters: less than 1 minute
- Choose Weight of Adapters: varies by AuthorMix method used (lorahub ~10min, sequential <1min)
- Obfuscation: varies by AuthorMix method used (lorahub ~5min, sequential 5min/style)
- Evaluation: ~10min
If you find this repository useful, or you use it in your research, please cite:
@misc{fisher2024styleremixinterpretableauthorshipobfuscation,
title={StyleRemix: Interpretable Authorship Obfuscation via Distillation and Perturbation of Style Elements},
author={Jillian Fisher and Skyler Hallinan and Ximing Lu and Mitchell Gordon and Zaid Harchaoui and Yejin Choi},
year={2024},
eprint={2408.15666},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2408.15666},
}
If you have any issues with the repository, questions about the paper, or anything else, please email [email protected] and [email protected].