Skip to main content

Your phone measures your daily habits.

Accelerometers, gyroscopes, microphones, magnetometers, barometers, and light sensors do more than orient a screen. The generated telemetry can distinguish hardware, classify posture and movement, recognize a gait, and turn repeated behavior into a map of places and routines.

sensor sampletime seriesinference
A phone turning sensor measurements into behavioral signalsAn angled phone surrounded by motion paths, audio waves, posture silhouettes, and labels for gait, room, and routine.xyzGAITROOMPOSTUREROUTINE
A measurement is small. A timestamped stream becomes a pattern. Repeated patterns become identity and context.

Sensor tracking has three different targets.

Hardware fingerprinting measures stable imperfections in the device itself. Behavioral biometrics models how a person walks, types, holds, tilts, or moves with the device. Context inference estimates what is happening around the device: whether it is on a bed, in a pocket, climbing floors, riding a train, or returning to the same room.

All three begin with ordinary measurements. The identifying information emerges from calibration bias, timing, repetition, and correlation with other records.

Try the motion demo ↓
one sensor streamx · y · z · time
01

Devicestable bias and manufacturing variation

02

Personcadence, handling, posture, gait

03

Place + routinerepeated context across time

The same measurements answer different questions depending on which features are retained and what other records are joined to them.

Watch posture inference change based on how you hold your phone.

Enable live sensors on a compatible phone, or run the simulation. This demo keeps samples inside the current tab and uses a deliberately simple classifier.

Waiting for a sensor stream

X
0.00
Y
0.00
Z
0.00
Samples0
Orientation estimateunknown
Motion estimateunknown

Production systems analyze windows rather than single samples. They extract frequency, variance, periodicity, transitions, and cross-sensor features, then associate the resulting labels with an account or device history.

Each channel describes a different part of the scene.

A single sensor is useful. Fused sensors reduce ambiguity: tilt plus angular velocity separates posture from movement; pressure adds floors; magnetic and acoustic signatures add place.

ACC01

Accelerometer

Measures
linear acceleration + gravity
Can reveal
posture, steps, gait, handling, vibration, calibration bias
GYR02

Gyroscope

Measures
angular velocity
Can reveal
rotation, gestures, gait, tap motion, low-frequency acoustic energy
MAG03

Magnetometer

Measures
magnetic field on three axes
Can reveal
heading, nearby structures, indoor magnetic signatures
BAR04

Barometer

Measures
air pressure
Can reveal
floor changes, elevation, weather-linked movement, travel mode
MIC05

Microphone + speaker

Measures
sound and frequency response
Can reveal
room acoustics, nearby media, ultrasonic beacons, hardware response
ENV06

Light + proximity

Measures
illumination and nearby surfaces
Can reveal
pocket state, indoor context, screen handling, recurring environments

“Lying down” can become “this is the bedroom.”

Accelerometer and gyroscope data can classify static postures and movement. A widely used UCI dataset labels walking, stairs, sitting, standing, and lying from phone motion. Add time, duration, recurrence, and location anchors, and a posture label becomes spatial context.

An inferred apartment map built from repeated sensor patternsA clear top-down apartment diagram follows a phone from morning movement in the kitchen, to a short evening recline in the living room, to a long overnight period lying still in the bedroom.LIVING ROOMBEDROOMKITCHENHALL12301UPRIGHT + STEPS07:04 · 18 MIN02SHORT RECLINE19:12 · 22 MIN03LYING + STILL23:41 · 7.2 HR · DAILYREPEATED DAILY ROUTE
Illustrative inference, not a literal floor-plan scan. Repeated posture, dwell time, movement paths, and environmental anchors can make probable rooms distinguishable.
  1. 01
    Classify posture

    Gravity across three axes separates upright, tilted, face-up, and side-on orientations.

  2. 02
    Measure dwell

    A few seconds is handling. Hours of repeated stillness at night is a routine.

  3. 03
    Anchor the pattern

    GPS, IP, Wi-Fi, Bluetooth, magnetic field, light, audio, or user activity can attach the routine to a place.

  4. 04
    Infer the room

    The longest recurring lying interval is a strong bed candidate; repeated paths and environmental signatures separate nearby spaces.

The phone can hear through several physical channels.

A speaker and microphone have a device-specific frequency response that can become a hardware fingerprint. A microphone can recognize room acoustics, nearby media, and ultrasonic markers. Even a MEMS gyroscope can react to low-frequency acoustic energy: the Gyrophone research recovered speaker information and partial speech features from gyroscope output.

Compare browser audio rendering →
frequency responseroom echonearby audioultrasonic ID
Different physical paths can identify the device, the surrounding space, or activity near it.

Sensor telemetry can expose hardware or behavior and drive inference.

01 · Hardware

Calibration fingerprint

Manufacturing and factory-calibration differences create stable offsets and scale factors. SensorID showed that calibration data could be recovered from motion output and used as a device fingerprint.

bias + scale + noise floor
02 · Person

Behavioral biometric

Walking cycles, tap force, typing rhythm, dominant-hand posture, and recurring gestures can form a model of the person carrying the phone.

cadence + timing + motion
03 · Context

Environmental signature

Pressure, magnetic anomalies, light, acoustics, motion transitions, and network anchors can distinguish floors, rooms, routes, and travel modes.

place + time + recurrence

The same hardware has several access points.

Access depends on platform, sensor, API, foreground state, sampling rate, and permission model. A visible prompt for one derived service does not describe every raw or fused sensor path.

SurfaceTypical reachIdentity attachment
Native application

Core Motion, Android Sensor Framework, microphone, platform SDKs

app account · device ID · telemetry session
Mobile website

DeviceMotion, DeviceOrientation, Generic Sensor APIs where supported

cookie · login · fingerprint · IP
Embedded SDK

analytics, fraud, advertising, engagement, crash and performance libraries

vendor IDs · publisher account · cross-app records
Operating system

activity recognition, fitness, safety, diagnostics, accessibility, system services

platform account · OS identifiers · service history

Permission is not a universal sensor boundary. Apple and Android expose different raw and derived services under different rules; browser support and prompts also vary. Sampling limits reduce resolution, but a lower-rate time series can still carry posture, gait, activity, and calibration features.

Control access, duration, and linkage separately.

No single switch covers native sensors, browser events, embedded SDKs, platform telemetry, and data already associated with an account.

  1. 01
    Review Motion, Fitness, Microphone, and Location access

    Remove access from apps that do not need the corresponding physical measurement.

  2. 02
    Limit background execution

    Short foreground samples and continuous histories expose different levels of routine.

  3. 03
    Separate identities and environments

    A sensor pattern becomes more useful when joined to a durable account, device ID, IP history, or precise location.

  4. 04
    Keep the sensor out of sensitive spaces

    For the strongest boundary, remove or power down the device rather than relying on an application-layer promise.