METHODOLOGY FOR DETECTING MALICIOUS CONTENT USING SENTIMENT ANALYSIS

Abstract

This paper investigates the problem of detecting malicious content in personal electronic communications, such as emails, social media messages, or messengers. Modern systems for filtering malicious content are mainly based on the analysis of keywords, structural features of messages, and technical parameters, such as IP addresses or message origin domains. Such systems have limited effectiveness against modern cyber threats that use message content with emotional manipulation or fake information. In addition, the effectiveness of such approaches is limited, especially when it comes to targeted attacks and new methods of bypassing protection mechanisms that are constantly evolving.

The paper proposes a hybrid model, "MaliciousContentDetector," which combines lexical methods, machine learning methods (TF-IDF, Random Forest), and deep learning (BERT) for text sentiment analysis. The model takes into account the linguistic features of the Ukrainian language and contextual emotional triggers. The proposed model is also implemented as a software which may be integrated into into browsers and other applications to enhance the cybersecurity of end-users. The developed model was tested on real messages from a popular news channel on a social network. During the analysis of 162 posts in that news channel, the "MaliciousContentDetector" model showed high efficiency and accuracy in detecting potentially malicious text content.