🕵 LLM Fingerprint

Identify LLMs by their response fingerprints

---

Research Question

Is it possible to identify which LLM generated a response by analyzing its semantic patterns across multiple standardized prompts?

Approach

LLM Fingerprint uses semantic similarity patterns across multiple prompts to create model-specific "fingerprints":

1. Fingerprint Creation: Generate multiple responses from known LLMs using standardized prompts

   • Fixed sampling parameters to ensure consistent sampling behavior
   • Multiple response samples per prompt to account for sampling variance
   • Several distinct prompts to capture model characteristics

2. Similarity Analysis: Measure semantic similarity within and between prompt response groups

   • Within-prompt similarity reveals consistency characteristics
   • Cross-prompt similarity patterns create a unique model signature

3. Model Identification: Match patterns from unknown models against the fingerprint database

   • Generate responses from the unknown model using the same standardized prompts
   • Compare similarity patterns with known models
   • Identify the closest matching fingerprint

Usage

Set required environment variables. See .envrc.example for more details.

Creating Model Fingerprints

# Generate samples for known model responses
llm-fingerprint generate \
  --language-model "model-1" "model-2" "model-3" \
  --prompts-path "./data/prompts/prompts_general_v1.jsonl" \
  --samples-path "samples.jsonl" \
  --samples-num 4

# Upload samples to ChromaDB
llm-fingerprint upload \
  --language-model "embedding-model" \
  --samples-path "samples.jsonl" \
  --collection-name "samples"

Identifying Unknown Models

# Generate samples for unknown model (or use an external service)
# Let's suppose the we don't know we are using model-2
llm-fingerprint generate \
  --language-model "model-2" \
  --prompts-path "./data/prompts/prompts_single_v1.jsonl" \
  --samples-path "unk-samples.jsonl" \
  --samples-num 1

# Query ChromaDB for model identification
llm-fingerprint query \
  --language-model "embedding-model" \
  --samples-path "unk-samples.jsonl" \
  --results-path "results.jsonl" \
  --results-num 2

# matches.jsonl will contain the results
# {"model": "model-2", "score": ... }
# {"model": "model-1", "score": ... }

Installation

The preferred way to install llm-fingerprint is using uv (although you can also use pip).

# Clone the repository
git clone https://github.com/S1M0N38/llm-fingerprint.git
cd llm-fingerprint

# Create a virtual environment
uv venv

# Install the package
uv sync # --all-groups # for installing ml and dev groups

Requirements

• Python 3.11+
• OpenAI-compatible API endpoints (/chat/completions and /embeddings)
• Access to ChromaDB (locally or hosted)

Contributing

This toy/research project is still in its early stages, and I welcome any feedback, suggestions, and contributions! If you're interested in discussing ideas or have questions about the approach, please start a conversation in GitHub Discussions.

For detailed information on setting up your development environment, understanding the project structure, and the contribution workflow, please refer to CONTRIBUTING.md.