🔐 SensitiveInfoDetector

A hybrid NLP system for detecting sensitive information (EMAIL + SECRET) using a fine-tuned DistilBERT model and regex rules.

---

🧠 Overview

SensitiveInfoDetector identifies sensitive entities such as emails and secrets (tokens / passwords) in free text.
It integrates a transformer-based model (DistilBERT) with rule-based pattern matching to improve precision and recall across different data distributions.

---

🖼️ App Previews

Example	Screenshot
✅ No Sensitive Information
⚠️ Warning
🚨 Critical
⚠️ + 🚨 Combined

---

🚀App Quick Launch (Colab / Local)

Clone the repository:

!git clone https://github.com/MaithaAlhammadi98/SensitiveInfoDetector.git
%cd SensitiveInfoDetector

Install dependencies:

!pip install -r requirements.txt

Launch the Gradio demo:

!python app/gradio_app.py

---

🧠 Dataset Details

This project uses a custom synthetic dataset designed for sensitive-information detection. All samples are anonymized and ethically generated.

Aspect	Description
Data Source	Synthetic / publicly available text snippets
Entities	EMAIL, SECRET (API keys, passwords, tokens)
Annotation	Manual labeling with Python span tagging
Dataset Size	≈ 1 500 train / 500 eval samples
Balance	Equal EMAIL and SECRET representation
Ethical Note	No real personal data used

🧩 The dataset trains the DistilBERT model to recognize and label sensitive entities within text data.

---

🧩 Architecture

Component	Description
DistilBERT	Fine-tuned transformer model for entity detection (EMAIL, SECRET).
Regex Rules	Deterministic patterns to catch edge cases missed by the model.
Hybrid Pipeline	Combines transformer predictions with regex results for higher recall.

---

🧪 Model Training

• Base model: distilbert-base-uncased
• Fine-tuning: Custom dataset with labeled EMAIL and SECRET entities.
• Framework: Hugging Face Transformers + PyTorch.
• Evaluation: Span-based precision, recall, and F1-score comparison across balanced and skewed datasets.

---

📊 Experimental Results

Dataset	Precision	Recall	F1-Score
Balanced	0.94	0.99	0.97
Skewed	0.97	0.98	0.98

Hybrid model consistently outperformed both the standalone transformer and rule-based baselines.

---

📈 Visual Results

Model	Confusion Matrix
DistilBERT
Rule-based

---

🌐 Online Resources

• GitHub Repository: MaithaAlhammadi98/SensitiveInfoDetector
• Hugging Face Model: Sensitive Info Detector (DistilBERT)

---

🧩 Folder Structure

SensitiveInfoDetector/
│
├── app/                 # Gradio interface
│   └── gradio_app.py
│
├── notebooks/           # Model training notebooks
│
├── evaluation/          # Metrics and confusion matrices
│   ├── span_metrics.py
│   ├── model_tp_fp_fn_bal.png
│
├── requirements.txt
└── README.md

---

🧑‍💻 Author

Maitha Alhammadi Master of Artificial Intelligence — University of Technology Sydney 📍 SensitiveInfoDetector is part of an NLP application-oriented project under Dr. Wei Liu.

---

🤖 AI Assistance Disclosure

ChatGPT (GPT-5, OpenAI) was used for debugging, report structuring, and documentation polishing in line with UTS academic integrity and ethical use guidelines.

---

📜 License

This project is licensed under the MIT License — free for academic and research use.

---