Adversarial Text Classifier

Prompt-injection detection using three classifiers: a TF-IDF logistic regression baseline, a pretrained DeBERTa-v3 model, and a fine-tuned version of it. Includes a FastAPI backend and React frontend for live inference.

Setup

From the repo root (requires uv):

uv sync

CUDA-enabled PyTorch is pinned in pyproject.toml. Use CPU wheels if you don't have an NVIDIA GPU.

Model checkpoints are not included in the repo. You need to either train them (see below) or obtain them separately and place them in Models/:

File	Description
Models/best_checkpoint_TF_IDF.pt	TF-IDF logistic regression weights
Models/tfidf_vectorizer.pkl	Fitted TF-IDF vectorizer
Models/finetuned_deberta-v2_best.pt	Fine-tuned DeBERTa weights

---

Notebooks

1. Pretrained DeBERTa — pretrained/pretrained_test.ipynb

The main experiment notebook. Covers:

• Loading the dataset (data/final_dataset/)
• Evaluating the pretrained protectai/deberta-v3-base-prompt-injection-v2 model
• Fine-tuning it on the combined dataset
• Gradient-based token attribution analysis

Run all cells in order. The fine-tuned checkpoint is saved to Models/finetuned_deberta-v2_best.pt.

2. TF-IDF Baseline — TF_IDF/TF_IDF_Training.ipynb

Trains a logistic regression classifier on TF-IDF features. Run all cells to produce:

• Models/best_checkpoint_TF_IDF.pt
• Models/tfidf_vectorizer.pkl

---

Backend (FastAPI)

Serves all three models (TF-IDF, pretrained DeBERTa, fine-tuned DeBERTa) with gradient-based explanations.

cd Backend
pip install -r requirements.txt
uvicorn main:app --reload

API runs at http://localhost:8000. Endpoints:

• GET /health — checks if models are loaded
• POST /predict — {"text": "your prompt here"}

Requires all model files in Models/ to be present before starting.

---

Frontend (React + Vite)

cd Frontend
npm install
npm run dev

Opens at http://localhost:5173. Make sure the backend is running first.