RiskRadar: an NLP-driven summarization system for query-based security insights

Abstract

The evolving complexity and frequency of cyber threat incidents demand the development of robust, user-friendly systems that can educate and assist users, and help them understand and mitigate them as much as possible. This thesis describes Risk Radar, a query-based information retrieval and response system that uses some advanced Natural Language Processing (NLP) methods to provide precise, contextaware responses to cybersecurity data. The system employs a multi-module architecture, with each module tailored to a specific task, such as query correction, semantic sequence analysis, information retrieval, and response generation. The core NLP models used are BERT for semantic similarity, BM25 for effective retrieval of relevant content, and a distilled BART model for summarization and context-based response generation. A unique rule-based mechanism improves query understanding and maintains contextual continuity across user interactions, addressing the challenges of multi-turn dialogue in technical. The proposed system not only provides detailed responses, but it also includes relevant articles to help users better understand specific incidents or trends. The system’s performance is measured by its ability to retain the context of user queries, retrieve and rank relevant content accurately, and generate coherent, informative responses. The system’s real-time implementation dynamically updates the dataset based on daily scraping of cybersecurity articles, ensuring that responses are timely and relevant. To address computational constraints, the model architecture prefers efficient methods like sequence-based rule application and DistilBART over more computationally intensive models like GPT-Neo. This trade-off balances accuracy and resource availability, resulting in a solution that is both practical and efficient. This thesis aims to contribute a scalable, efficient solution for tackling the growing need for real-time, user-oriented cybersecurity information systems.