NewsGuard-LLM: A Lightweight, Dynamic Large Language Model based Framework for Real-Time Threat Detection in Digital News

The accelerated evolution of global digital news ecosystems, driven by increased connectivity and the proliferation of networked platforms, has placed digital news as a high-volume open-source asset for threat intelligence. This development has intensified the need for real-time threat detection in critical sectors, enabling organizations to anticipate and mitigate emerging risks more effectively. While Large Language Models (LLMs) offer advanced capabilities in natural language understanding and content analysis, their widespread operational deployment remains limited by substantial computational demands, reliance on external cloud infrastructures, and growing concerns over data privacy and governance. To this end, this paper presents a lightweight and scalable LLM framework designed for real-time threat detection in digital news headlines. Unlike traditional LLM architectures, the proposed system operates entirely within the organization’s infrastructure, ensuring data privacy and minimizing external resource dependencies. The framework automates end-to-end pipeline, from data collection and labeling to dynamic retraining and deployment, leveraging GPU acceleration to enhance computational efficiency. The system integrates open-source LLMs for initial data annotation, a dynamic feedback loop for continuous model refinement, and a user-friendly dashboard for real-time threat monitoring. The power and cost analysis of the framework demonstrate that GPU-based training, while drawing higher instantaneous power (433W), completes tasks up to four times faster than CPU-based training, resulting in lower overall energy consumption (0.108 kWh per training run) and reduced cost per task ($0.0141 vs. $0.0166). The experimental evaluation of the proposed framework shows the classification accuracy of 96% and 70% reduction in training time.