Skip to main content
Be Bitwise

Device Fingerprinting

Mobile Privacy & Data Protection · Mobile Advanced Security track

10 min read

In this module

Beyond Identifiers: Fingerprinting

Cookies got blocked. Advertising identifiers got walled off. The tracking industry pivoted to device fingerprinting — identifying a device by what it looks like rather than what it stores. Fingerprinting is harder to spot, harder to stop, and there's nothing to reset.

What Makes a Device Unique

Every phone carries a stew of hardware and software quirks that, taken together, draw a near-unique silhouette. No single attribute outs you on its own. Stack a few dozen and you separate one device from millions with high confidence.

Each attribute carries a different amount of entropy. Device model is low-entropy — millions share the same iPhone. Combine device model + OS build + timezone + language + installed fonts + screen brightness + accessibility settings, and you're closing in on a single human.

Passive Fingerprinting

Passive fingerprinting just listens. It collects what the device emits during normal operation — no probes, no active code, no fuss.

User Agent String

The HTTP User-Agent header gives away the browser, OS, and device. On mobile, that means model, OS version, browser version. The string alone doesn't pin you down, but it shrinks the population you're hiding in by orders of magnitude.

Screen and Display

Screen resolution, pixel density, colour depth, and even the presence of a notch or Dynamic Island (visible through CSS safe areas) all feed the fingerprint. On iOS, the right combination of resolution and safe-area insets is enough to identify the exact iPhone model.

Installed Fonts

Devices ship with different font sets. Add user-installed fonts and the entropy climbs further. Web fingerprinters probe which fonts are available by rendering text in each and measuring the resulting pixel dimensions.

Sensor Data

Research has shown that sensor readings contain device-specific signatures:

  • Accelerometer calibration: Manufacturing variations in the accelerometer's zero-g offset and sensitivity produce device-unique readings. Even two phones of the same model produce slightly different accelerometer values when at rest.
  • Gyroscope bias: Similar to accelerometers, gyroscopes have manufacturing-induced biases that are stable over time and unique per device.
  • Battery status: The combination of battery level, charging state, and estimated time remaining (available through the Battery Status API in browsers) was shown to be surprisingly unique, enabling short-term tracking.
Sensor Calibration Fingerprinting

Sensor calibration fingerprinting exploits the fact that MEMS sensors (accelerometers, gyroscopes) have manufacturing-induced variations in their calibration parameters (offset, gain, cross-axis sensitivity). These variations are stable over time and unique per device. Research has demonstrated that accelerometer readings alone can distinguish individual devices with over 95% accuracy. The defence is challenging because the sensor APIs are legitimately needed for screen rotation, gaming, and fitness applications.

Active Fingerprinting

Active fingerprinting runs code that pokes the device's rendering engine, audio stack, or WebGL implementation, then reads back the device-specific output.

Canvas Fingerprinting

Canvas fingerprinting is the workhorse of active fingerprinting, deployed across an absurd portion of the web. It rides on tiny differences in how devices rasterise graphics:

  1. A JavaScript script creates a hidden HTML5 <canvas> element.
  2. The script draws specific text strings (using various fonts) and geometric shapes.
  3. The script calls toDataURL() to extract the rendered pixels as a data URI.
  4. The data URI is hashed to produce a fingerprint.

Pixels diverge across devices for boring reasons: GPU silicon, graphics driver version, font renderer (ClearType on Windows, Core Text on macOS/iOS, FreeType on Android), anti-aliasing algorithm, sub-pixel rendering. Together they're plenty.

Canvas fingerprints are stable across sessions on the same device and different across distinct devices. Surveys have found canvas fingerprinting scripts on a frankly alarming share of the top websites.

WebGL Fingerprinting

WebGL fingerprinting extends canvas fingerprinting to 3D rendering:

  • The script queries WEBGL_debug_renderer_info to obtain the GPU vendor and renderer strings.
  • The script renders a 3D scene and hashes the output pixels.
  • WebGL parameter queries (maximum texture size, shader precision, supported extensions) add further entropy.

The GPU renderer string plus rendering output is brutally discriminating, because driver implementations all produce slightly different pixels for the same OpenGL/WebGL commands.

Audio Fingerprinting

Audio fingerprinting exploits differences in how devices process audio through the Web Audio API:

  1. A script creates an AudioContext and generates a signal (typically an oscillator).
  2. The signal is processed through a chain of audio nodes (compressor, analyser).
  3. The processed audio samples are extracted and hashed.

Devices process the same audio slightly differently because the DAC (Digital-to-Analogue Converter), driver, and OS audio stack all introduce micro-variations. Audio fingerprinting is less famous than canvas, but it adds entropy you can combine with the rest.

Native App Signals

Most of the techniques above come from browser research, but an installed app has a much richer menu, and that's where most mobile fingerprinting happens:

  • Android: Settings.Secure.ANDROID_ID (since Android 8 scoped to the app's signing key, user and device, but stable until factory reset), Build properties such as model, manufacturer and fingerprint, and the list of other installed packages. Android 11's package visibility rules mean apps must declare which other apps they want to see, and Google Play restricts the broad QUERY_ALL_PACKAGES permission.
  • iOS: the IDFV (identifier for vendor) is shared by all apps from one developer on a device, with no ATT prompt needed. Keychain items can outlive an uninstall, so an app can recognise a reinstall. DeviceCheck gives developers two bits per device, stored by Apple and surviving reinstalls, and App Attest proves a genuine app instance. Both are meant for fraud prevention, but DeviceCheck's bits in particular let a developer recognise a device it has seen before.
  • Both: locale, time zone, storage size, boot time and similar system values. Apple's privacy manifests now require apps to give an approved reason for reading several of these, because they're classic fingerprinting inputs.

Battery Status API

The Battery Status API (navigator.getBattery()) was identified as a fingerprinting vector by researchers in 2015. The API returns:

  • Battery level (0.0 to 1.0, with platform-dependent precision).
  • Whether the device is charging.
  • Estimated time to full charge.
  • Estimated time to discharge.

Take those four values at one moment in time and the combination is shockingly unique. Enough to track someone across sites within a single browsing session, on a window of minutes to hours.

Firefox pulled the Battery Status API from web pages in 2017 (Firefox 52), citing exactly this abuse. Safari never shipped it. Chrome still exposes it, though only to secure (HTTPS) pages. The specification now lets browsers reduce the precision of the values, and the original 2015 study's sharpest result came from implementations reporting the level with far more digits than anyone needed.

Platform Anti-Fingerprinting Measures

iOS

Apple has shipped a stack of anti-fingerprinting countermeasures:

  • Intelligent Tracking Prevention (ITP) in Safari limits cookie lifetimes and cross-site tracking.
  • Safari does not expose the Battery Status API.
  • Since Safari 17, Private Browsing adds small amounts of noise to canvas and Web Audio readback, so the same page gets slightly different results each session. Ordinary browsing gets less protection.
  • Sensor data access requires user permission in web contexts.
  • The User-Agent string carries less than it used to: it has never named the specific iPhone model, and Safari has frozen parts of it, such as the OS version, so it changes less between devices.

Android

Google's anti-fingerprinting efforts focus on:

  • Chrome's User-Agent reduction, which freezes most of the UA string and moves detail into Client Hints that sites have to request.
  • Chrome limits the Battery Status API to secure contexts.
  • Android 10+ randomises MAC addresses for WiFi probe requests.
  • Sensor access from background apps is restricted.

Limitations

Anti-fingerprinting runs into a tough problem: most of the inputs trackers exploit are also useful for legitimate work. Screen resolution drives responsive design. Font lists matter for text rendering. GPU capabilities are how WebGL apps decide what to ship. Lock these APIs down and you break honest developers while determined trackers route around the restriction with new signals.

Research Frontiers

Sensor Calibration Fingerprinting

Academic work has shown accelerometer and gyroscope data identifying individual devices with 95%+ accuracy, though under lab conditions and across sets of tens to a few hundred devices, not a general rate for the whole population. The 2019 SensorID attack went further, recovering the factory calibration of iPhone motion sensors from a web page, and Apple responded in iOS 12.2 by adding noise to the sensor output. Mitigation is awkward — these sensors are core to screen rotation, games, and health tracking. The most viable defence so far: inject noise into sensor readings below human perception but above fingerprinting precision.

Accelerometer Uniqueness

Researchers have shown that the tiny vibrations produced by a device's own motors — haptic engine, vibration motor — are measurable through its accelerometer and vary device to device. That's a hardware-level fingerprint with no software knob to turn.

Key Concept
Device Fingerprinting

Device fingerprinting identifies devices by combining hardware and software characteristics into a near-unique signature. Passive fingerprinting collects attributes transmitted during normal operation (user agent, screen size, sensor calibration). Active fingerprinting probes the rendering engine (canvas, WebGL) and audio subsystem to extract device-specific outputs. Platforms have implemented countermeasures (Safari noise injection, Battery API removal, MAC randomisation) but face a fundamental tension: the data points used for fingerprinting serve legitimate purposes, making complete prevention difficult without breaking functionality.