Abstract
Fake News (FN) prevalence on Online Social Networks (OSNs) and online websites is a worldwide issue. The previous studies on Fake News Detection (FND) have focused on rich-resource languages with limited relevance to users other than native speakers. Despite the progressing multilingual approaches, FND in low-resource languages remains obscure due to a lack of large-sized annotated corpora of real-world news. Large Language Models (LLMs) have emerged as a promising solution for multilingual FND. This study leverages the power of LLMs for an automated mechanism compared to traditional feature extraction methods. We curate the first large-sized multi-domain corpus, the Hook and Bait Urdu, with 78,409 fake and true news, and fine-tune the LLaMA 2 model for our proposed approach. We implement the curated dataset for two experiments. First, this study evaluates the dataset for unimodal text-based Urdu FND. The proposed LLaMA-based approach shows an accuracy of 0.978 and an F1-score of 0.971. For our second experiment, we fine-tuned the LLaMA 2-based framework for multilingual FND using the curated dataset (in Urdu) and the ISOT Fake News dataset (in English). Analytical and prediction performance comparisons with the previous studies validate the efficacy of the proposed framework with an accuracy of 0.984 and an F1-score of 0.980. The lightweight LoRA fine-tuning method, with 0.032% trainable parameters, ensured robust data handling, computational efficiency while leveraging early stopping and optimized hyperparameters for reliable and high-performing monolingual and multilingual FND. The real-world news dataset is publicly available for developing an automated FND mechanism to curb the threat of FN and related cybercrimes.