Predictive analysis of telecommunication network fraud crimes based on big language modeling

This paper proposes a telecom fraud crime analysis framework that integrates privacy protection, domain knowledge injection, parameter-efficient fine-tuning, Prompt engineering, and time-series prediction. A dataset of 500 real cases is collected from a police platform, covering typical fraud types such as rebate fraud, impersonating law enforcement, and false investments. First, a two-stage de-identification process using regular expressions and BERT-NER is applied to thoroughly de-sensitize sensitive information, ensuring data compliance while retaining essential case elements. Next, a knowledge base of over 200 police-specific terms and common knowledge is constructed, and the MEMIT method is used to locally inject police terminology into the LLaMA model, significantly enhancing the model’s understanding of terms such as “running points” and “card farmers.” The experimental results show that, after knowledge editing, entity extraction precision improved from 77.3% to 85.0%, with recall increasing by 6.4% points. Further fine-tuning with LoRA improved precision and recall by 3.3 and 3.9% points, respectively. Finally, the case classification Macro-F1 score reached 0.862, outperforming TextCNN (0.846) and BERT fine-tuning models (0.850). This framework demonstrates strong performance in telecom fraud case analysis and provides valuable support for intelligent policing and crime prediction.