An Information Reliability Framework for Detecting Misinformation based on Large Language Models

Information plays a key role in decision making, influencing others and risk assessment. However, malicious actors often falsify facts to mislead viewers. With the advent of Generative AI models, creating and modifying information has become much easier and more accessible, significantly reducing the time and effort required for attackers. To address the challenges of misinformation and disinformation, this paper presents the "Information Reliability Framework (IRF)," a detection framework built with DarkBERT (a dark web-focused language model) [1] and LLaMA-3.1 (a large language model), which incorporates real-time web data for validation. DarkBERT was fine-tuned using the FEVER dataset (Fact Extraction and VERification) [2] to enhance its efficiency in misinformation, disinformation, and fact-checking tasks. This fine-tuning enables DarkBERT to assign a reliability score to a given statement by analyzing malicious keywords and contexts, which is then combined with LLaMA-3.1 to generate a detailed explanation for the user, leveraging its knowledge and logical reasoning. To minimize reliance on outdated information and reduce hallucinations in LLaMA-3.1, the framework integrates real-time web data using Retrieval-Augmented Generation (RAG). The framework was evaluated alongside standalone LLaMA-3.1 with RAG using two synthetic datasets: generic statements and DarkBERT-specific statements, generated by ChatGPT-4o with its latest "search the web" feature. The evaluation showed that the IRF provided promising results across metrics such as accuracy, precision, recall, and F1 score, making it effective for validating text from various sources, including social media, emails, articles, messages, websites, and blogs.