Institutional Repository of Vilnius University
Grėsmių socialinėse medijose stebėjimo įrankis /
Abstract
dc:descriptionA bachelor’s thesis presents a social media threat monitoring information system intended for organizations’ threat intelligence and security analysts. The study found that assessing subjective messages using only isolated risk indicators can be inaccurate, since ambiguous phrasing often lacks clear, explicit signals. Therefore, message assessment was computerized by applying an integrated, multi aspect evaluation and a unified risk scoring approach, enabling more consistent prioritization of content based on a defined threshold. A competitor analysis showed that most market solutions focus on technical cybersecurity scenarios, while analysis of communication content related to intent interpretation is typically offered only as an auxiliary feature. The implemented system integrates data collection, natural language processing, threat profile application, risk assessment, and incident management. It collects messages from Reddit, MockPost, and Discord, stores them in a database, applies NLP analysis, and performs intent classification, topic detection, sentiment analysis, and named entity recognition. The system uses configurable threat profiles with keyword sets and a weighted evaluation model, which enables automatic incident creation when the risk score exceeds the defined threshold. A management dashboard provides analysts with incident review, filtering, commenting, and data export capabilities. Testing indicated that intent detection quality depends primarily on a purpose built and curated training dataset. Model comparison showed that context aware language models better capture ambiguous formulations, and DistilBERT provides a favorable balance between accuracy and performance, therefore it was selected as the system’s primary model. Key conclusions: 1. Assessing threats in subjective content requires an integrated, context based evaluation that reduces interpretation errors and supports consistent message prioritization. 2. A three tier architecture and a unified data structure enable consistent processing of messages from multiple platforms using the same analytical principles. 3. The implemented system automates message collection, analysis, and incident creation, while the dashboard supports review workflows and decision documentation. 4. The best intent detection performance was achieved using DistilBERT, while further accuracy improvements depend on expanding the dataset and adding multilingual support. Keywords: threat monitoring, social media, natural language processing, intent detection, risk assessment, incident management, OSINT, DistilBERT.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Gricuk, Ričardas,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD277688838&prefLang=en_US