BotSim: LLM-Powered Malicious Social Botnet Simulation

Boyu Qiao, Kun Li, Wei Zhou, Shilong Li, Qianqian Lu, Songlin Hu

TL;DR

This paper introduces BotSim, a novel LLM-powered simulation framework for creating realistic malicious social botnets. They leverage this framework to generate BotSim-24, a new dataset of highly human-like LLM-driven bots interacting on Reddit, demonstrating that existing bot detection methods struggle significantly against these advanced adversaries, highlighting the urgent need for new detection strategies.

What Problem Does This Paper Address?

Online social networks (OSNs) like X (formerly Twitter) and Reddit are crucial for global communication, but they are increasingly threatened by sophisticated social bots. While traditional bots were often rule-based, the advent of Large Language Models (LLMs) has enabled bots to exhibit unprecedented intelligence, mimicking human profiles, conversations, and interactions with remarkable accuracy. This makes them highly effective at spreading misinformation and poses a significant challenge for platform regulation and cybersecurity. Existing bot detection methods, largely developed for simpler bots and trained on older datasets, are proving inadequate against these new LLM-powered threats. Furthermore, creating realistic datasets of these advanced bots is difficult due to ethical constraints and the challenges of manual annotation when bots are so human-like.

Methods

The core contribution is the BotSim framework, a scalable simulation environment designed to model the behavior of LLM-driven malicious social bots. BotSim creates a virtual social environment populated by both real human accounts and intelligent agent bots powered by LLMs. The framework includes:

  • Social Environment: Mimics real OSN dynamics, including account collection, message feeding, timeline setup, and interaction modes.
  • Environment Perception: Allows agent bots to "perceive" the social environment, including account profiles, message streams, and interaction data, by using structured prompts for LLMs.
  • Action List: A set of common social media actions (e.g., Create User, Post, Comment, Like, Repost) that agent bots can execute.
  • Agent Decision Center: The brain of the bot, integrating goal tasks, role settings, background knowledge, environmental perception, action lists, and memory to make decisions and plan actions.

Using this framework, the authors generated the BotSim-24 dataset. This dataset is unique because it includes both real Reddit users and LLM-driven agent bots that are designed to be highly human-like. They employed disguise techniques based on metadata, textual content, and interaction patterns to make the bots indistinguishable from real users. The LLM (GPT40-mini) was used to generate bot profiles, posts, and comments, mimicking human behavior and interaction styles.

Datasets

  • Real OSN Data: Six popular news-related SubReddits on Reddit ("worldnews", "politics", "news", "InternationalNews", "UpliftingNews", "GlobalTalk") were used to construct the social environment. Data was collected from June 20, 2023, to June 19, 2024.
  • Background Knowledge Data: News articles from authoritative sources (BBC, NBC News, NYTimes, People's Daily) and fact-checking sites (Truthorfiction, Snopes) were collected to provide context for bot disinformation campaigns.
  • BotSim-24 Dataset: The primary contribution, this dataset was generated using the BotSim framework. It comprises 1,907 human accounts and 1,000 LLM-driven agent bot accounts, simulating interactions across six SubReddits. This dataset is notable for its focus on LLM-driven bots and rich interaction data.
  • Existing Datasets for Benchmarking: Cresci-15, Cresci-17, TwiBot-20, TwiBot-22, and MGTAB-22 were used to compare the performance of detection methods on BotSim-24 against traditional bot datasets.

Key Insights and Ideas

  • LLM Bots are a New Frontier: The intelligence and human-like capabilities of LLM-driven bots represent a significant leap in sophistication, rendering many existing detection methods obsolete.
  • The Power of Simulation: BotSim provides a scalable and adaptable way to generate realistic bot behavior and create up-to-date datasets, crucial for staying ahead of evolving bot threats.
  • Human-like Camouflage is Key: The success of LLM bots lies in their ability to mimic human metadata, text, and interaction patterns, making them incredibly difficult to distinguish.
  • Graph-based Methods Show Promise, but with a Caveat: While graph neural networks (GNNs) generally outperform other methods, their effectiveness on BotSim-24 is partly due to an "incomplete" graph structure (lack of human-to-bot directed edges), which paradoxically helps GNNs identify differences. This suggests that truly human-like bot interactions might further challenge GNNs.
  • Human Annotators Struggle Too: Even human annotators find it challenging to accurately identify LLM-driven bots, underscoring the difficulty of the problem.

Key Results

  • BotSim-24 Challenges Existing Methods: Detection methods that perform well on traditional datasets like Cresci-15 and TwiBot-20 showed significantly poorer performance on BotSim-24. For instance, some methods performed almost like random guessing.
  • Metadata and Textual Disguise are Effective: Traditional machine learning approaches relying on metadata were undermined by the bots' effective camouflage. Text-based methods also struggled due to LLMs' advanced text comprehension and generation capabilities.
  • Heterogeneous Graph-based Methods Perform Best: GNNs, particularly heterogeneous graph approaches like RGCN and S-HGN, showed superior performance compared to feature-based and homogeneous graph methods on BotSim-24.
  • Edge Perturbation Experiment: Simulating varying levels of human-to-bot interactions showed that as these interactions increase, detection performance initially declines, highlighting the vulnerability of current methods to more complex bot behaviors.
  • LLM Detector Performance: Table 5 shows that even LLM-based detectors struggle with text-based detection, with accuracy often below 60% for some models, though GPT-4-turbo-ca showed promising few-shot learning capabilities.

Limitations and Future Work

  • Cost of LLMs: The authors acknowledge that the cost of using LLMs limited the scale of the BotSim-24 dataset.
  • Lack of Human-Bot Interactions: The simulation environment, while incorporating human data, doesn't fully capture real-time human-initiated interactions towards bots, creating an incomplete graph structure.
  • Future Work:
    • Developing more comprehensive and larger-scale bot detection datasets.
  • Engaging domain experts to better simulate real user behavior.
  • Crowdsourcing to create more extensive virtual simulation environments with real human and bot interactions.
  • Further research into GNNs and other methods to better handle the nuanced interactions of advanced LLM-driven bots.

Summary updated

Abstract

Social media platforms like X(Twitter) and Reddit are vital to global communication. However, advancements in Large Language Model (LLM) technology give rise to social media bots with unprecedented intelligence. These bots adeptly simulate human profiles, conversations, and interactions, disseminating large amounts of false information and posing significant challenges to platform regulation. To better understand and counter these threats, we innovatively design BotSim, a malicious social botnet simulation powered by LLM. BotSim mimics the information dissemination patterns of real-world social networks, creating a virtual environment composed of intelligent agent bots and real human users. In the temporal simulation constructed by BotSim, these advanced agent bots autonomously engage in social interactions such as posting and commenting, effectively modeling scenarios of information flow and user interaction. Building on the BotSim framework, we construct a highly human-like, LLM-driven bot dataset called BotSim-24 and benchmark multiple bot detection strategies against it. The experimental results indicate that detection methods effective on traditional bot datasets perform worse on BotSim-24, highlighting the urgent need for new detection strategies to address the cybersecurity threats posed by these advanced bots.
BibTeX
@inproceedings{qiao2025botsim,
  author = {Boyu Qiao and Kun Li and Wei Zhou and Shilong Li and Qianqian Lu and Songlin Hu},
  title = {{BotSim: LLM-Powered Malicious Social Botnet Simulation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  url = {https://openreview.net/forum?id=uWECMKk3kM},
}
Figure