Robust Fake News Detection using Large Language Models under Adversarial Sentiment Attacks

Technical Summary: Robust Fake News Detection using Large Language Models under Adversarial Sentiment Attacks

Problem Statement
Misinformation and fake news pose a significant societal challenge, necessitating reliable automated detection methods. While prior research has established sentiment as a critical signal for distinguishing fake from real news, this reliance creates a vulnerability. Adversaries can manipulate the emotional tone of text to evade detection, a risk exacerbated by the advent of Large Language Models (LLMs) capable of rewriting content with controlled sentiment while preserving factual accuracy. Although some studies have explored adversarial attacks on fake news detectors, they have primarily focused on stylistic features (e.g., writing style) rather than sentiment. Consequently, the specific vulnerability of fake news detectors to LLM-generated sentiment manipulation remains largely unexplored.

Methodology
The paper proposes AdSent, a sentiment-robust detection framework designed to maintain consistent veracity predictions across both original and sentiment-altered news articles. The framework consists of two primary components:

1. Sentiment-Targeted Adversarial Benchmarking:

   • The authors utilize LLMs as "counterfeiters" to generate controlled adversarial attacks.
   • For each news document, the LLM rewrites the text to convey positive, negative, or neutral sentiment while strictly preserving factual content and coherence.
   • This process creates a benchmark to evaluate the robustness of state-of-the-art detectors against sentiment shifts, treating detectors as black-box classifiers.

2. Sentiment-Agnostic Training Strategy:

   • To mitigate the bias introduced by sentiment cues, the authors propose a training strategy where the detection model is fine-tuned on sentiment-neutralized variants of the training data.
   • An LLM (specifically LLaMA-3.1-8B-Instruct) is used to rewrite training articles into a neutral tone.
   • The detection model is then trained on these neutralized versions, forcing it to rely on factual content rather than emotional tone to determine veracity.
   • During inference, input articles are first neutralized by the counterfeiter before being passed to the detector.

Key Contributions

1. Adversarial Benchmarking: The authors introduce a suite of controlled sentiment-based attacks using LLMs. They demonstrate that current state-of-the-art detectors suffer significant performance degradation when faced with sentiment manipulation.
2. Discovery of Model Bias: The study reveals a critical bias in existing models: they tend to classify neutral-toned articles as "real" and non-neutral (emotional) articles as "fake." This finding contradicts the initial hypothesis that fake news is inherently more negative; instead, the manipulation of sentiment to a neutral tone proves to be the most challenging adversarial scenario.
3. AdSent Framework: The paper introduces AdSent, a novel approach that enhances robustness by training on sentiment-neutralized data. This ensures prediction consistency regardless of the article's emotional framing.

Experimental Results
The authors evaluated AdSent on three benchmark datasets: PolitiFact, GossipCop, and Labeled Unreliable News (LUN).

• Vulnerability of Baselines: Under sentiment manipulation, standard detectors (including fine-tuned RoBERTa and SheepDog) experienced substantial performance drops, with F1-scores decreasing by up to 21.51% on the PolitiFact dataset. Zero-shot LLMs also showed significant degradation.
• Impact of Sentiment Shifts: Analysis of prediction flips indicated that converting fake news to a neutral tone caused the highest rate of misclassification (false negatives), where fake news was incorrectly labeled as real. Conversely, real news was less likely to be misclassified as fake when neutralized.
• Performance of AdSent: AdSent significantly outperformed all baselines on neutralized test sets, achieving macro F1 scores of 87.76 on PolitiFact and 78.56 on GossipCop.
• Generalization:
  • Cross-Attack: AdSent demonstrated robustness not only against sentiment attacks but also generalized effectively to style-based adversarial attacks (e.g., mimicking specific publishers), outperforming the specialized SheepDog baseline in those scenarios.
  • Cross-Dataset: When trained on PolitiFact and tested on the unseen LUN dataset, AdSent maintained superior performance compared to SheepDog, indicating that adversarial sentiment training yields transferable features.

Significance and Claims
The paper claims that the reliance of current fake news detectors on sentiment cues is a critical weakness that LLMs can easily exploit. By demonstrating that neutral sentiment is the most effective adversarial perturbation, the authors highlight a fundamental bias in existing models. The significance of AdSent lies in its ability to decouple veracity detection from emotional tone, thereby creating a more robust defense against sophisticated, LLM-driven misinformation campaigns. The authors conclude that their approach significantly improves both accuracy and robustness, generalizing well to unseen datasets and diverse adversarial scenarios.

Future Directions
The authors note that future work will focus on extending AdSent to multimodal misinformation detection, investigating how visual signals (e.g., emotional expressions in images) interact with textual sentiment. They also plan to explore adversarial perturbations targeting other journalistic news values, such as proximity, novelty, and prominence.