GPU timing fingerprint (DrawnApart)
Times many identical GPU draw calls to expose per-chip speed differences.
In one line
Can time many GPU drawing jobs to tell apart even identical graphics chips
Times many identical GPU draw calls to expose per-chip speed differences.
Manufacturing variation in execution units makes even identical GPU models slightly different. DrawnApart separated identical GPUs after only 150 ms of drawing and remained stable for weeks.
What it can and can't do
What the user does: Nothing. It happens when the page loads
It can
- Run many identical drawing tasks and time each one
- Spot tiny speed differences between two chips of the same model
- Separate identical GPUs after only about 150 ms of drawing
- Recognise the same device again on later visits
It cannot
- Show anything on your screen
- Work if graphics tests are blocked
- Ask permission first
- Know who you are by itself
Browser support
| Browser | Status | First version | Released | Note |
|---|---|---|---|---|
| Chrome | Supported | 9 | 2011-02-03 | None |
| Edge | Supported | 79 | 2020-01-15 | None |
| Firefox | Supported | 4 | 2011-03-22 | None |
| Safari | Supported | 8 | 2014-10-16 | None |
| Safari on iPhone | Supported | 8 | 2014-09-17 | None |
| Chrome on Android | Supported | 9 | Not listed | None |
| Samsung Internet | Supported | 5.0 | 2016-12-15 | None |
How a site asks for it
| API | WebGL drawArrays loops; performance.now(); WebGPU compute |
| How access is granted | none |
| What the user sees | Not listed |
| Needs a secure context (HTTPS) | No |
| Needs a user gesture | No |
| How long it lasts | Not listed |
| How to take it back | Not listed |
| Inside an iframe | Not listed |
Typical flow
- Run vertex-heavy draw loops
- Time each pass with high resolution timers
- Classify with a trained model
Data you get
What comes back: Vector of per-pass timings
Shape: hash
| Field | Type | Unit | Range | Precision |
|---|---|---|---|---|
duration | float | ms | 0.1 to 100 | timer precision limited |
Technical limits
- Needs repeated GPU load and good timers
- Research prototype, not known in mass deployment
The story
What it is
A script times many identical GPU drawing jobs to expose tiny speed differences between individual graphics chips.
For you
You never see it. Your graphics chip does a brief busy job, and how long each part takes is a little different from the next chip, even of the same model. It acts as a hardware level ID.
For makers
Researchers showed this can tell apart two identical phones or laptops. It matters because it keeps working when other signals are hidden. Ad and fraud tools may adopt it, so browsers are working to blur timers.
Where you see it
| Type | Example | Apps |
|---|---|---|
| security | A fraud engine tries to recognize one specific laptop even if its software looks brand new. | None named |
| ads | An ad tracker could separate two people who both own the same new iPhone model. | None named |
| games | A browser benchmark times your GPU to pick settings, and it is the same measure that can leak identity. | None named |
| analytics | A research team tests how many devices can be told apart by their graphics timing alone. | DrawnApart (research) |
| creative | A web 3D modeling app times its draw calls to adjust detail so the view stays smooth. | None named |
| banking | A bank looks for stable hardware traits so a new login on a stolen password stands out. | None named |
| developer | A browser maker adds timer rounding so these tricks give less precise results. | None named |
Risks
- Hard to hide hardware traits
- Mostly in research so far
Good uses and abuse
Legitimate uses
- None
Known abuse
- Per-device tracking beyond GPU model
Scores
| Score | Value | Meaning |
|---|---|---|
| Sensitivity | 4 of 5 | How revealing the data is |
| Utility | 1 of 5 | How useful it is to honest apps |
| Friction | 0 of 5 | How much the user has to do |
| Fingerprint bits | 6 | Rough identifying entropy, when known |
History
- 2022-02, paper: Laor et al. NDSS 2022 present DrawnApart; improve tracking duration 67% over software fingerprints
Related capabilities
Related
- WebGL render hash. Can draw a hidden 3D scene and use the pixels to tell your device apart
- High resolution timer precision. Can use a clock whose sharpness shows your browser and privacy mode
Sources
How well checked: Checked against real behavior.