🧠 MIA-Spector: Membership Inference Analysis Platform

MIA-Spector is a unified platform designed for privacy risk evaluation on Large Language Models (LLMs), providing Membership Inference Attack (MIA) analysis and visualization.

The platform contains two major components:

1. 🧩 MIA-Inspector API (Backend) — FastAPI-powered inference & decision service
2. 💡 MIA-Portal (Frontend) — React + Tailwind interactive visualization console

---

🌟 Features Overview

Module	Function	Tech Highlights
Backend — MIA-Inspector API	Unified LLM loading (Pythia, LLaMA, etc.) with MIA metric computation and decision	✔ Async FastAPI ✔ Auto model registry + caching ✔ Prometheus performance metrics
Frontend — MIA-Portal	Interactive dashboard for selecting model/config and performing single-sample MIA decision	✔ Modern dark UI ✔ Responsive layout ✔ Pretty JSON rendering
Security Layer (Auth + Rate Limit)	API key authentication with Token-Bucket rate control	✔ Dynamic .env loading ✔ Configurable QPS limit ✔ Per-client counters
Metric Engine	Supports Min-K%, Min-K++, PPL, Renyi entropy, etc.	✔ YAML-based threshold loading ✔ Multiple inference modes (Youden J / FPR@α)
System Monitoring	/metrics endpoint export	✔ Traffic stats, latency histograms, GPU monitoring

---

🧩 Project Structure

MIA-Spector/
│
├── service/
│   ├── app/              ← Backend API (FastAPI)
│   │   ├── main.py       # Entrypoint
│   │   ├── deps.py       # Model loading & caching
│   │   ├── middlewares/  # Auth + rate limit
│   │   ├── routers/      # Health, Meta, Decision
│   │   └── config.py     # Global config (MODELS, CFGS)
│   │
│   ├── portal/           ← Frontend (React + Vite + Tailwind)
│   │   ├── src/
│   │   │   ├── pages/Console.jsx
│   │   │   ├── components/SectionCard.jsx
│   │   │   └── index.css
│   │   └── vite.config.js
│   │
│   └── uvicorn.run.sh    # One-click launch script
│
├── attacks/               # Core MIA metric algorithms
├── src/                   # Utility modules
├── configs/               # YAML thresholds
├── models/                # Local HF model weights
└── README.md

---

⚙️ Installation

🖥️ Windows & Docker Support

For detailed instructions on running MIA-Spector on Windows 11 (via Docker or Native), please refer to WINDOWS.md.

Backend Environment (Linux/Mac)

conda create -n mia-inspector python=3.11
conda activate mia-inspector
pip install -r requirements.txt

Frontend

cd service/portal
npm install

## 🚀 Launch & Usage

### 🔹 Start Backend

```bash
cd service/app
bash ../uvicorn.run.sh

or manually:

uvicorn app.main:app --host 0.0.0.0 --port 8080 --reload

.env Example

API_KEYS=abc123,def456
REQUIRE_AUTH=True

Authentication Test

curl -H "Authorization: Bearer abc123" http://localhost:8080/healthz

---

🔹 Start Frontend

cd service/portal
npm run dev

Visit: http://localhost:5173

Initial Setup on UI:

1. Set API Base URL: http://localhost:8080
2. Enter API Key (e.g., abc123)
3. Click Load Models / Configs
4. Input text → Run decision

---

🎯 Backend API Endpoints

Path	Method	Description
/healthz	GET	Health check
/v1/models	GET	Available model registry
/v1/configs	GET	Threshold configs
/v1/decide	POST	Membership inference decision
/metrics	GET	Prometheus monitoring export

Example request:

{
  "text": "The mitochondrion is the powerhouse of the cell.",
  "family": "pythia",
  "model": "pythia-410m",
  "cfg": "WikiMIA_length128",
  "metric_group": "mink++",
  "subkey": "0.3",
  "mode": "bestJ"
}

---

💡 Frontend Highlight Features

Module	Function	Tech
Config Panel	Bind API base, key, clientId	React Hooks + LocalStorage
Model Browser	Fetch /v1/models	Axios + JSON viewer
Config Browser	Fetch /v1/configs	Responsive grid, dark-mode tuned
Inference Panel	Text input + model/metric selection	Tailwind form components
Result Viewer	Pretty JSON rendering	font-mono + shadow cards
Theme	Dark-mode optimized	dark:bg-slate-900 + custom gray scales

---

🧠 Background

Membership Inference Attack (MIA) evaluates whether a model leaks training samples by observing its output distribution (e.g., PPL, Min-K%, confidence gaps).

MIA-Spector provides:

• Single-sample decision + batch analysis
• Multi-metric fusion & directional decision
• ROC / AUC visualization
• Cross-model threshold sharing via YAML configs

---

📊 Example Output

{
  "decision": "Uncertain",
  "confidence": 0.67,
  "score": -0.76,
  "threshold": -0.91,
  "metric_group": "mink++",
  "subkey": "0.3",
  "direction": "+",
  "mode": "bestJ"
}

---

🧾 Citation

@misc{MIA-Spector2025,
  title  = {MIA-Spector: Unified Platform for Text and Image Membership Inference Analysis},
  author = {Liu, Jiajun and Collaborators},
  year   = {2025},
  url    = {https://github.com/JiajunLiu/MIA-Spector}
}

⚠️ This project is intended only for privacy & security research and defensive analysis. Any misuse for unauthorized attacks or data leakage is strictly prohibited.