Data and Code for "Fake News in Sheep's Clothing: Robust Fake News Detection Against LLM-Empowered Style Attacks" (KDD 2024)

This repo contains the data and code for the following paper:

Jiaying Wu, Jiafeng Guo, Bryan Hooi. Fake News in Sheep's Clothing: Robust Fake News Detection Against LLM-Empowered Style Attacks, ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD) 2024.

Abstract

It is commonly perceived that fake news and real news exhibit distinct writing styles, such as the use of sensationalist versus objective language. However, we emphasize that style-related features can also be exploited for style-based attacks. Notably, the advent of powerful Large Language Models (LLMs) has empowered malicious actors to mimic the style of trustworthy news sources, doing so swiftly, cost-effectively, and at scale. Our analysis reveals that LLM-camouflaged fake news content significantly undermines the effectiveness of state-of-the-art text-based detectors (up to 38% decrease in F1 Score), implying a severe vulnerability to stylistic variations. To address this, we introduce SheepDog, a style-robust fake news detector that prioritizes content over style in determining news veracity. SheepDog achieves this resilience through (1) LLM-empowered news reframings that inject style diversity into the training process by customizing articles to match different styles; (2) a style-agnostic training scheme that ensures consistent veracity predictions across style-diverse reframings; and (3) content-focused veracity attributions that distill content-centric guidelines from LLMs for debunking fake news, offering supplementary cues and potential interpretability that assist veracity prediction. Extensive experiments on three real-world benchmarks demonstrate SheepDog's style robustness and adaptability to various backbones.

Requirements

python==3.7.0
numpy==1.22.4
torch==1.10.0+cu111
transformers==4.13.0

Data

Our work is based on the PolitiFact and GossipCop datasets from the FakeNewsNet benchmark, and the LUN dataset from (Rashkin et al., 2017).

We provide the data files utilized for training and evaluating SheepDog under data/. In our data files, the label 0 represents real news, and the label 1 represents fake news.

Original Unaltered Training / Test Articles

The .pkl files under data/news_articles/ contain the unaltered news article texts. Please refer to Section 6.1.1 of our paper for more details.

Adversarial Test Sets

The .pkl files under data/adversarial_test/ contain the four adversarial test sets under LLM-empowered style attacks, denoted as A through D. Please refer to Section 4.1 and Table 4 of our paper for a detailed formulation.

LLM-Empowered News Reframings

The .pkl files under data/reframings/ contain the style-diverse reframings of training articles. Please refer to Section 5.1 of our paper for detailed descriptions.

Content-Focused Veracity Attributions

The .pkl files under data/veracity_attributions/ contain the content-focused veracity attributions of training articles. Here, each article is assigned 4 binary labels according to the following 4 attributions: (1) lack of credible sources, (2) false or misleading information, (3) biased opinion, and (4) inconsistencies with reputable sources. Please refer to Section 5.3 of our paper for detailed descriptions.

Run SheepDog

Start training with the following command:

sh train.sh

Model checkpoints will be saved under checkpoints/, and results will be saved under logs/.

Additionally, under logs/logs_archive_all4_adv/, we provide archived experiment logs for SheepDog on both the original test set and adversarial test sets A-D.

Contact

jiayingwu [at] u.nus.edu

Citation

If you find this repo or our work useful for your research, please consider citing our paper

@inproceedings{wu2024sheepdog,
author = {Wu, Jiaying and Guo, Jiafeng and Hooi, Bryan},
title = {Fake News in Sheep's Clothing: Robust Fake News Detection Against LLM-Empowered Style Attacks},
year = {2024},
booktitle = {Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
pages = {3367–3378}
}