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

New Framework Enhances Fake News Detection Amid Sentiment Manipulation

Research has unveiled a new framework, AdSent, designed to bolster the effectiveness of fake news detection mechanisms against sentiment manipulation tactics. This development responds to the increasing sophistication of misinformation strategies that employ large language models (LLMs) to alter sentiment in news articles.

Prior studies have established sentiment as a vital indicator for identifying fake news, but this reliance exposes vulnerabilities, as adversaries can exploit sentiment cues to bypass detection systems. While some research has examined adversarial samples generated by LLMs, the emphasis has primarily been on stylistic elements rather than sentiment manipulation.

AdSent Framework Overview

• Controlled Sentiment-Based Adversarial Attacks: AdSent generates adversarial samples that specifically target sentiment alterations, providing insights into how sentiment shifts affect detection performance.
• Impact Analysis: Modifications in sentiment significantly influence the performance of fake news detection systems, with neutral articles more frequently classified as real, while non-neutral sentiments are often identified as fake.
• Sentiment-Agnostic Training Strategy: AdSent employs a training strategy that minimizes the influence of sentiment on detection outcomes.

Performance and Generalization

Extensive experiments demonstrate that AdSent surpasses existing competitive baselines in accuracy and improves robustness, effectively generalizing across unseen datasets and various adversarial scenarios.