Device Identification via Canvas Fingerprinting (Variational Inference)

Probabilistic model for identifying device identity from JavaScript canvas fingerprint data using variational inference.

Overview

This project explores whether a device can be identified just from the way it renders images.
Given high-dimensional RGBA pixel data generated from deterministic browser rendering, we model the device identity as a latent variable and infer:

P (Image | Device)

The goal is to reverse the rendering process probabilistically and determine which device most likely produced a given image.

Key Features

• Variational inference framework for device identification
• High-dimensional RGBA pixel input from canvas rendering
• Latent variable modeling of device identity
• Designed to handle cross-device rendering variability and noise

Methodology

1. Collect labeled canvas fingerprint data (image → device)
2. Represent each image as a flattened RGBA pixel vector
3. Train a probabilistic model where:
   • Observed variable: image data
   • Latent variable: device identity
4. Perform inference to predict device given new image input

Results

• Demonstrates feasibility of probabilistic device identification from rendering artifacts
• Captures uncertainty in predictions across similar device fingerprints
• Highlights challenges in distinguishing devices with near-identical outputs

Instructions

• Run "pip install -r requirements.txt"
• To run the frontend collection website run:

npm install -g firebase-tools
firebase login
firebase serve

• To run the variational inference model

python train.py

• The backend is hosted on firebase so not available immediately from the repo