Reproduce “Fake News in Sheep's Clothing: Robust Fake News Detection Against LLM-Empowered Style Attacks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiaying Wu, Jiafeng Guo, Bryan Hooi
Researchers at National University of Singapore and University of Chinese Academy of Sciences introduce SheepDog, a style-robust fake news detector that leverages LLM-empowered reframings and a multi-task learning paradigm to prioritize content over style. SheepDog consistently improved F1 scores by 2.59% to 15.70% on adversarial test sets, effectively countering LLM-empowered style attacks against existing detectors which saw performance drops of up to 38%.
Key insights
- The reliance on stylistic cues in fake news detection creates a severe vulnerability against sophisticated LLM-driven style manipulation.
- Training models to be invariant to stylistic variations by exposing them to diverse stylistic reframings of the same content is crucial for robustness.
- Combining fine-tuned LMs with the reasoning capabilities of LLMs for data augmentation and attribution guidance can build more resilient and interpretable detection systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.