SCIENCE@home

An LLM-based Behavior Modeling Framework for Malicious User Detection

Meng Jiang,Wenjie Wang,Chongming Gao,Shaofeng Hu,Kaishen Ou,Hui Lin,Fuli Feng

Published 2025 in International Conference on Information and Knowledge Management

ABSTRACT

Malicious users pose significant threats to social platforms. Extensive efforts have leveraged user behavior sequences to model relationships between various actions and capture behavioral patterns for malicious user detection; however, they rely on behavior IDs, ignoring valuable behavior content such as self-introductions in friend requests, which offer crucial clues for detecting malicious user. We thus propose leveraging Large Language Models (LLMs) to jointly model IDs and content in user behavior sequences. The key to effective malicious user detection is to infer malicious user behavior patterns. However, inferring these patterns from labeled behavior sequences suffers from poor data efficiency and limited generalization, resulting in suboptimal malicious user detection performance. To overcome the limitations, we propose leveraging the malicious user specifications (e.g., definitions and common deceptive tactics) from the existing expert handbook. These specifications guide LLMs in reasoning over user behavior IDs and content before making predictions. To this end, we introduce an LLM-based behavior modeling framework with an expert handbook to enhance LLMs' behavior reasoning. We first distill the user's behaviors into a concise summary, guided by malicious user specifications in the expert handbook, and then feed the summary and users' demographic features into LLMs for comprehensive reasoning and detection. We conduct extensive online and offline experiments on the Weixin platform, validating the superiority of the proposed framework over the original Weixin detection baseline, achieving, for example, a 5.34% improvement in F1-Score.

PUBLICATION RECORD

  • Publication year

    2025

  • Venue

    International Conference on Information and Knowledge Management

  • Publication date

    2025-11-10

  • Fields of study

    Computer Science

  • Identifiers
    DOI 10.1145/3746252.3761520
  • External record

    Open on Semantic Scholar

  • Source metadata

    Semantic Scholar

CITATION MAP

EXTRACTION MAP

CLAIMS

  • No claims are published for this paper.

CONCEPTS

  • No concepts are published for this paper.

REFERENCES

Showing 1-40 of 40 references · Page 1 of 1

CITED BY

  • No citing papers are available for this paper.

Showing 0-0 of 0 citing papers · Page 1 of 1