waf-distilbert

WAF-DistilBERT: Web Application Firewall using DistilBERT

Model Description

WAF-DistilBERT is a fine-tuned version of DistilBERT, specifically trained to detect malicious web requests in real-time. This model serves as the core component of a Web Application Firewall (WAF) system.

Intended Use

This model is designed for:

• Real-time detection of malicious web requests
• Integration into web application security systems
• Identifying common web attacks like SQL injection, XSS, and path traversal
• Enhancing existing security infrastructure

Out-of-Scope Use Cases

This model should not be used as:

• The sole security measure for web applications
• A replacement for traditional WAF rule-based systems
• A tool for generating malicious payloads
• A security measure for non-HTTP traffic

Training Data

The model was trained on the CSIC 2010 HTTP Dataset, which includes:

• Normal HTTP requests
• Various attack patterns including SQL injection, XSS, buffer overflow
• A balanced distribution of benign and malicious requests

Training Procedure

• Base model: DistilBERT-base-uncased
• Training type: Fine-tuning
• Training hardware: NVIDIA GPU
• Number of epochs: 3
• Batch size: 32
• Learning rate: 2e-5
• Optimizer: AdamW
• Loss function: Binary Cross-Entropy

Performance and Limitations

Performance Metrics

• Accuracy: >95%
• F1-Score: >0.94
• False Positive Rate: <1%
• Average inference time: <100ms per request

Limitations

• Limited to HTTP request analysis
• May require retraining for organization-specific traffic patterns
• Performance may vary for zero-day attacks
• Best used in conjunction with traditional security measures

Bias and Risks

Bias

The model may show bias towards:

• Common attack patterns in the training data
• English-language payloads
• HTTP requests following standard web frameworks

Risks

• False positives may block legitimate traffic
• False negatives could allow attacks through
• May require regular updates to maintain effectiveness
• Resource consumption under high load

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("jacpacd/waf-distilbert")
model = AutoModelForSequenceClassification.from_pretrained("jacpacd/waf-distilbert")

# Prepare input
request = "GET /admin?id=1 OR 1=1"
inputs = tokenizer(request, return_tensors="pt", truncation=True, max_length=512)

# Make prediction
with torch.no_grad():
    outputs = model(**inputs)
    prediction = torch.sigmoid(outputs.logits)

is_malicious = prediction.item() > 0.5
confidence = prediction.item()

Environmental Impact

• Model Size: ~268MB
• Inference Energy Cost: Low (compared to larger models)
• Training Energy Cost: Moderate

Technical Specifications

• Model Architecture: DistilBERT
• Language(s): English
• License: MIT
• Input Format: Text (HTTP requests)
• Output Format: Binary classification with confidence score
• Model Size: 268MB
• Number of Parameters: ~65M

Citation

If you use this model in your research, please cite:

@misc{waf-distilbert,
  author = {jacpacd},
  title = {WAF-DistilBERT: Web Application Firewall using DistilBERT},
  year = {2025},
  publisher = {Hugging Face},
  journal = {Hugging Face model repository},
  howpublished = {\url{https://huggingface.co/jacpacd/waf-distilbert}}
}

Contact

For questions and feedback about the model, please:

• Open an issue on GitHub
• Contact through Hugging Face
• Submit pull requests for improvements