CSE-ARS: Deep Learning-Based Late Fusion of Multimodal Information for Chat-Based Social Engineering Attack Recognition

Abstract

With the increasing prevalence of chat-based social engineering (CSE) attacks targeting unsuspecting users, the need for robust defenses has never been more critical. In this paper, we introduce Chat-based Social Engineering Attack Recognition System (CSE-ARS), an innovative and effective CSE defense system. CSE-ARS employs a late fusion strategy that integrates the findings of five specialized deep learning models, each focused on detecting distinct CSE attack enablers: critical information leakage recognizer (CRINL-R), personality traits recognizer (PERST-R), dialogue acts recognizer (DIACT-R), persuasion recognizer (PERSU-R), persistence recognizer (PERSI-R). The system harnesses weighted linear aggregation and employs simulated annealing with 10-fold cross-validation, ensuring optimal model performance. CSE-ARS is trained on the CSE-ARS Corpus, a carefully curated dataset tailored to the intricacies of CSE attacks. Extensive evaluation reveals that CSE-ARS achieves satisfactory results in identifying and neutralizing CSE threats, enhancing user security in online interactions.