LLM Fingerprinting System

A black-box fingerprinting system that identifies the underlying LLM model family (GPT, LLaMA, Mistral, etc.) by analyzing response patterns across 75 discriminative prompts. The system can identify fine-tuned models as well, tracing them back to their foundational base model.

Note: Check config.py to see all identifiable model families

You can find an already NLP trained model in the model directory.

Supported Backends

Backend	Description	API Key Required
ollama	Local Ollama instance	❌ No
ollama-cloud	Ollama Cloud API	✅ OLLAMA_CLOUD_API_KEY
openai	OpenAI API (or compatible)	✅ OPENAI_API_KEY
gemini	Gemini API (or compatible)	✅ GEMINI_API_KEY
deepseek	Deepseek API (or compatible)	✅ DEEPSEEK_API_KEY
custom	Custom HTTP request	✅ CUSTOM_API_KEY

Installation

pip install -r requirements.txt

# Or install as a package
pip3 install -e .

# Optional: Download NLTK data for text processing
python -c "import nltk; nltk.download('punkt_tab'); nltk.download('stopwords')"

Quick Start

Ollama

# Identify model and fine-tuning

llm-fingerprinter identify -b ollama --model some-model 

# Train your own classifier
# Fingerprint the LLM
llm-fingerprinter simulate --model llama3.2 --family llama
# Train on the Fingerprints
llm-fingerprinter train

Custom - Interact with any LLM via HTTP request

llm-fingerprinter identify -r ./custom_request.txt --api-key <API_KEY>
# Example of custom request inside the example folder

Ollama Cloud

export OLLAMA_CLOUD_API_KEY="your-key"
llm-fingerprinter simulate -b ollama-cloud --model llama3.2 --family llama

OpenAI

export OPENAI_API_KEY="your-key"
llm-fingerprinter simulate -b openai --model gpt-4 --family gpt

Gemini

export GEMINI_API_KEY="your-key"
llm-fingerprinter simulate -b gemini --model gemini-2.5-pro --family gpt

Deepseek

export DEEPSEEK_API_KEY="your-key"
llm-fingerprinter simulate -b deepseek --model deepseek-v3.2 --family deepseek

Custom API

export CUSTOM_API_KEY="your-key"
llm-fingerprinter simulate -b custom -e http://your-api.com/v1 --model your-model --family llama

---

Commands

Backend Options (all LLM commands)

Option	Short	Default	Description
--backend	-b	custom	Backend: ollama, ollama-cloud, openai,deepseek,gemini ,custom
--endpoint	-e	auto	API endpoint URL
--api-key	-k	env var	API key

simulate

Run fingerprinting simulations for training data.

llm-fingerprinter simulate [OPTIONS]

Option	Default	Description
--model	required	Model name
--family	required	Family: gpt, claude, llama, gemini, mistral, qwen, gemma
--num-sims	optional	Number of simulations
--repeats	optional	Prompt repeats per simulation

Examples:

# Ollama local
llm-fingerprinter simulate --model llama3.2 --family llama

# Ollama Cloud
llm-fingerprinter simulate -b ollama-cloud --model llama3.2 --family llama

# OpenAI
llm-fingerprinter simulate -b openai --model gpt-4 --family gpt --num-sims 5

# Custom endpoint
llm-fingerprinter simulate -b openai -e https://api.groq.com/openai/v1 -k $GROQ_KEY --model llama-3.1-70b --family llama

train

Train classifier from saved fingerprints.

llm-fingerprinter train [--augment/--no-augment]

identify

Identify model family using trained classifier.

llm-fingerprinter identify --model <model-name> [-b <backend>]

Other commands

list-models

List available models on the API.

llm-fingerprinter list-models [-b <backend>]

list-fingerprints

List saved fingerprints by family.

llm-fingerprinter list-fingerprints

info

Show configuration and status.

llm-fingerprinter info

---

Environment Variables

Variable	Backend	Description
OLLAMA_CLOUD_API_KEY	ollama-cloud	Ollama Cloud API key
OPENAI_API_KEY	openai	OpenAI API key
GEMINI_API_KEY	gemini	Gemini API key
DEEPSEEK_API_KEY	deepseek	DeepSeek API key
CUSTOM_API_KEY	custom	Custom API key
LOG_LEVEL	all	Logging level (DEBUG, INFO, etc.)

How It Works

1. 75 Prompts across 3 layers:

   • Stylistic: Analyze writing style and formatting preferences
   • Behavioral: Assess response patterns and decision-making behavior
   • Discriminative: Identify model-specific characteristics and inconsistencies

2. Feature Extraction: 384-dim embeddings + 12 linguistic + 6 behavioral features

3. PCA reduction to 64 dimensions (Optional)

4. Ensemble Classification: Random Forest (45%) + SVM (45%) + MLP (10%)

---

Contributing

Contributions are welcome! Whether you're adding support for new models, improving accuracy, or extending to additional clients, please see CONTRIBUTING.md for guidelines.

---

License

MIT License