Social Media Messages for Early Cyberattack Detection on Blockchain
홈페이지 · 논문 1편
# ELTEX-Blockchain: A Domain-Specific Dataset for Cybersecurity
🔐 12k Synthetic Social Media Messages for Early Cyberattack Detection on Blockchain
## Dataset Statistics
| Category | Samples | Description |
|--------------|-------------|-----------------|
| Cyberattack | 6,941 | Early warning signals and indicators of cyberattacks |
| General | 4,507 | Regular blockchain discussions (non-security related) |
## Dataset Structure
Each entry in the dataset contains:
- message_id: Unique identifier for each message
- content: The text content of the social media message
- topic: Classification label ("cyberattack" or "general")
## Performance
Gemma-2b-it fine-tuned on this dataset:
- Achieves a Brier score of 0.16 using only synthetic data in our social media threat detection task
- Shows competitive performance on this specific task when compared to general-purpose models known for capabilities in cybersecurity tasks, like granite-3.2-2b-instruct, and cybersecurity-focused LLMs trained on Primus
- Demonstrates promising results for smaller models on this specific task, with our best hybrid model achieving an F1 score of 0.81 on our blockchain threat detection test set, though GPT-4o maintains superior overall accuracy (0.84) and calibration (Brier 0.10)
## Attack Type Distribution
| Attack Vectors | Seed Examples |
|-------------------|------------------------|
| Social Engineering & Phishing | Credential theft, wallet phishing |
| Smart Contract Exploits | Token claim vulnerabilities, flash loans |
| Exchange Security Breaches | Hot wallet compromises, key theft |
| DeFi Protocol Attacks | Liquidity pool manipulation, bridge exploits |
For more about Cyberattack Vectors, read Attack Vectors Wiki
## Citation
``bibtex
@misc{razmyslovich2025eltexframeworkdomaindrivensynthetic,
title={ELTEX: A Framework for Domain-Driven Synthetic Data Generation},
author={Arina Razmyslovich and Kseniia Murasheva and Sofia Sedlova and Julien Capitaine and Eugene Dmitriev},
year={2025},
eprint={2503.15055},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2503.15055},
}
``