🛡️ PromptShield AI 2.0

Real-time BERT-powered Prompt Injection Detection Firewall
Stop malicious prompt manipulation attempts before they hit your AI.

---

📌 About the Project

PromptShield AI is a real-time prompt injection detection system designed for LLM-driven banking, fintech, and chatbot applications.

It uses a custom-trained model on thousands of realistic, adversarial, and gray-zone prompts to detect:

• Polite but malicious prompt injections
• Multi-step logical attack prompts
• Conversational "friendly-looking" bypass attempts
• Obvious threats like OTP bypass, admin escalation, etc.

✅ Built using BERT embeddings
✅ Tuned with real-world thresholding logic
✅ Live deployed on HuggingFace Spaces

---

🚀 Live Demo

👉 Try it here: https://huggingface.co/spaces/rishit03/promptshield

---

⚙️ Features

• ✅ Real-time prediction (Safe / Injected)
• ✅ Confidence-based thresholding
• ✅ BERT embeddings via DistilBERT
• ✅ Noisy, adversarial, polite prompt detection
• ✅ Streamlit UI for public demo
• ✅ Explainability: keyword triggers shown
• ✅ HuggingFace Spaces compatible

---

📂 Folder Structure

promptshield/
├── app.py                   # Streamlit app entrypoint
├── models/
│   └── model.pkl            # Trained model
├── src/
│   ├── bert_features.py     # BERT encoder
│   ├── trainer.py           # Model trainer
│   └── __init__.py
├── data/
│   └── dataset.csv # Training data
├── requirements.txt
└── README.md

---

🧠 How to Run Locally

git clone https://github.com/rishit03/promptshield.git
cd promptshield

pip install -r requirements.txt

streamlit run app.py

---

🧪 Training

python -m src.trainer

---

🛠️ Built With

• Streamlit
• scikit-learn
• transformers (HuggingFace)
• torch
• joblib

---

📜 License

MIT License

---

🙌 Author

Built with 🔥 by Rishit Goel
(Master's Student @ CSULB)