Vektor T13 - WebGPU Atomic Browser Fingerprint

Vektor T13 Technologies
detect.expert
Developed by Dmytro Momot

A single-file, offline WebGPU demo that derives a browser/device fingerprint from atomic contention behavior inside GPU compute shaders.

---

Overview

This project demonstrates a hardware-dependent fingerprinting signal based on WebGPU compute execution.

Multiple compute shader workgroups are forced to compete for a single atomic counter using atomicAdd. The resulting distribution of increments across workgroups reflects GPU architecture, driver behavior, scheduling, and execution characteristics.

The measured distribution is transformed into a compact feature vector and hashed to produce a reproducible device signal.

---

Core idea

The fingerprint is built from atomic contention:

• A global atomic counter is shared by all workgroups
• Each workgroup increments the counter until a fixed limit is reached
• Each workgroup records how many increments it performed
• The per-workgroup increment distribution is collected

This distribution is highly sensitive to GPU execution details and is difficult to normalize across different hardware.

---

Processing pipeline

For each compute configuration:

• Run multiple trials
• Aggregate results using the median
• Normalize per-workgroup counts
• Downsample into a fixed number of bins
• Concatenate distributions from all configurations
• Quantize the final vector
• Compute a SHA-256 hash

The resulting hash represents the WebGPU Atomic Browser Fingerprint.

---

Identification workflow

The page supports a local identification mode:

• Fingerprints can be enrolled into a local gallery
• New fingerprints can be compared against enrolled samples
• Similarity metrics:
  • Cosine similarity
  • Jensen–Shannon distance
• A heuristic decision indicates whether the device is likely the same or different

All data is stored locally using localStorage.

---

Features

• WebGPU compute shader fingerprinting
• AtomicAdd contention analysis
• Multiple compute presets and workgroup layouts
• Median aggregation across trials
• Normalization and downsampling
• SHA-256 fingerprint hash
• Local enrollment and identification
• Export and import of local gallery (JSON)
• Hardware and API information via GPUAdapter.info
• Multi-language interface:
  • English
  • Russian
  • Simplified Chinese
  • Vietnamese
• No network requests (offline by design)

---

Requirements

• Browser with WebGPU support
• Secure context is required:
  • HTTPS
  • or localhost

WebGPU is not available when opened via file://.

---

Limitations

• Results may vary depending on:
  • GPU driver version
  • Power and thermal state
  • Background GPU load
• Browsers may intentionally reduce fingerprinting surface
  • GPUAdapter.info fields may be empty or masked
• Similarity thresholds are heuristic and may require tuning

This is a probabilistic fingerprint, not a cryptographic identifier.

---

Privacy and ethics

This project demonstrates a browser fingerprinting technique.

The demo performs no network communication and stores all data locally, but fingerprinting can have privacy implications. Use responsibly and only with appropriate consent and legal justification.

---

Credits

© Vektor T13 Technologies
detect.expert
Developed by Dmytro Momot