Fraudulent Message Detection Using a BERT-Based Semantic Model Integrated with a Rule Engine

View at publisher Summarise with VIRAT
Authors: Boyuan Ba, Miao Zang, Jie Yang, Chunqin Huang
Cited by: 0
References: 7

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

The rapid proliferation of fraudulent messages poses significant challenges to digital communication security. To address this, we propose a hybrid detection framework that synergistically integrates BERT-based semantic representation with a rule engine for fine-grained fraud identification. During the optimization phase, BERT is fine-tuned on an annotated dataset of fraudulent and legitimate texts using a maximum sequence length of 128 tokens, a batch size of 16, a learning rate of 2e–5, and three training epochs. The training process adopts the AdamW (Adaptive Moment Estimation with Weight Decay) optimizer together with a linear warm-up schedule, enabling the model to better learn contextual dependencies and subtle semantic variations. Initially, BERT is fine-tuned on annotated datasets comprising fraudulent and legitimate texts, enabling the model to capture latent semantic structures and contextual dependencies. Subsequently, a rule engine leveraging sensitive word lists, format constraints, and scenario templates is employed to reinforce the detection of specific fraud patterns. The sensitive-word lexicon is constructed from public fraud-alert cases, domain-specific terminology, and high-frequency expressions identified in real-world datasets. These lexical items are then organized into category-oriented vocabularies—such as impersonation, refund scams, investment schemes, and policy-driven fraud—each accompanied by structural constraints and scenario-based regular expression templates to capture typical linguistic patterns used by scammers. Empirical evaluation on the validation set demonstrates that the proposed hybrid model achieves over 95% accuracy, 95% recall for known fraud patterns, and maintains a false-positive rate below 5%. Compared to standalone BERT or rule-based approaches, our method significantly enhances both detection precision and computational efficiency. This framework provides a robust and practical solution for real-time monitoring and mitigation of fraudulent messages, offering valuable insights for the development of automated fraud detection systems in diverse communication environments.

Fields of Study & Subjects

Spam and Phishing Detection Misinformation and Its Impacts Hate Speech and Cyberbullying Detection

Publication details

Journal / source
International Conference on Artificial Intelligence and Social Network Systems (AISNS)
Pages
13–17
Publication date
9 January 2026
Language
EN