This model is a fine-tuned version of bert-base-multilingual-uncased for detecting website defacement attacks using textual content. It is designed for binary classification of web pages into legitimate or defaced.
The model was developed in the context of cybersecurity for Industry 4.0, focusing on automated detection of defacement incidents.
Intended uses & limitations
Intended uses
Web monitoring systems
Cybersecurity pipelines
Automated detection of defaced pages
Limitations
Uses only textual content (no images, scripts, or structure)
May miss attacks located at the end of pages (truncation to 512 tokens)
Performance evaluated on a balanced dataset (may differ in real-world scenarios)
Dataset is predominantly in English
Training data
The model was trained on a balanced dataset of 98,086 web pages, composed of:
Defaced pages from Zone-H
Legitimate pages from TRANCO
Each sample contains the textual content of a webpage.
The dataset is not publicly available due to privacy and ethical concerns .
Training procedure
Base model: bert-base-multilingual-uncased
Task: text classification
Learning rate: 2e-5
Batch size: 32
Early stopping applied
Maximum sequence length: 512 tokens (head-only truncation)
The dataset includes content from defaced websites, which may contain sensitive or offensive material. For this reason, the dataset is not publicly distributed.
Use of this model should follow responsible cybersecurity and data privacy practices.
Citation
If you use this model, please cite:
Romão, H. L. (2026).
Detecção de Ataques de Desfiguração em Páginas Web Utilizando Transferência de Aprendizagem e Transformers.
Universidade Federal de Roraima.
