LLM-GAN: Construct Generative Adversarial Network Through Large Language Models For Explainable Fake News Detection

TL;DR

The paper tackles explainable fake news detection by addressing two core challenges: LLMs can be misled by plausible fake news, and explanations can be unreliable. It introduces LLM-GAN, a prompting-based framework in which an LLM acts as both Generator and Detector, trained through inter-adversary prompting and refined via self-reflection prompting to produce accurate predictions with human-friendly explanations. The approach leverages real-news augmentation, adversarial data generation, and autonomous self-improvement, achieving state-of-the-art performance on Weibo21 and GossipCop while delivering higher-quality explanations as assessed by GPT-4o. The authors further demonstrate practical impact by integrating LLM-GAN into a cloud-native platform, enabling scalable, explainable fake news detection in real-world settings.

Abstract

Explainable fake news detection predicts the authenticity of news items with annotated explanations. Today, Large Language Models (LLMs) are known for their powerful natural language understanding and explanation generation abilities. However, presenting LLMs for explainable fake news detection remains two main challenges. Firstly, fake news appears reasonable and could easily mislead LLMs, leaving them unable to understand the complex news-faking process. Secondly, utilizing LLMs for this task would generate both correct and incorrect explanations, which necessitates abundant labor in the loop. In this paper, we propose LLM-GAN, a novel framework that utilizes prompting mechanisms to enable an LLM to become Generator and Detector and for realistic fake news generation and detection. Our results demonstrate LLM-GAN's effectiveness in both prediction performance and explanation quality. We further showcase the integration of LLM-GAN to a cloud-native AI platform to provide better fake news detection service in the cloud.

Figures (6)

• Figure 1: Comparisons of performance in fake news detection on the Weibo21 dataset. The simply prompted LLM (orange) underperforms compared to existing deep-learning-based methods, especially when predicting fake news. For more details, please refer to Sec. \ref{['sec:experiments']}.
• Figure 2: The architecture of our proposed LLM-GAN model for explainable fake news detection. LLM-GAN consists of two main stages: (i) inter-adversary prompting that allows the Detector to benefit from the news-faking process; and (ii) self-reflection prompting that automates the Detector to revise itself from its past mistakes.
• Figure 3: Illustrative output examples of LLM-GAN for two cases where the strongest baseline model, ARG ARG, failed. LLM-GAN makes correct predictions and provides logical, human-readable explanations as evidence. Errors in fake news and key information in explanations are highlighted with a gray background.
• Figure 4: Ablation study for inter-adversary prompting mechanism with the varying number of input collected news items. LLM-GAN can effectively benefit from the news-faking process via data augmentation of the input news.
• Figure 5: Comparisons of industrial application between LLM-GAN and baselines. Our LLM-GAN can significantly boost user trust in the system.