SecureSense separates the trusted embedded decision path from the optional operator interface.
The Nano 33 BLE Sense Lite samples its onboard LSM9DS1 and APDS9960 sensors at 20 Hz. Every 20 readings form a non-overlapping one-second window. Firmware calculates 45 deterministic features and evaluates:
- a 100-tree binary random forest for
normalversussuspicious; - an experimental 23-node decision tree for event subtype;
- a five-window majority vote for output stability;
- a training-derived light-range safeguard for extreme enclosure-opening changes.
The built-in LED and USB Serial JSON are controlled by the device decision.
In device mode, the Flask application does not load or execute an ML model.
It reads device JSON, displays telemetry and stores a capped local audit trail
in SQLite.
replay and serial-host modes do run host inference, but the interface labels
that boundary explicitly.
When explicitly enabled in device mode, the console forwards only stable
security-state transitions to an ntfy topic. Publishing runs in a background
thread and cannot block serial acquisition. Repeated suspicious states are
suppressed by a cooldown, and raw IMU, light and proximity telemetry are not
included.
This gateway is not part of the inference path. If the Internet, Mac, dashboard or third-party notification service is unavailable, the Arduino model, LED and serial output continue to operate locally.
| Component | Signals | Purpose |
|---|---|---|
| LSM9DS1 | x/y/z acceleration, magnitude | movement and vibration context |
| APDS9960 | clear light, proximity | enclosure and near-device interaction |
Temperature, camera and microphone data are outside the deployed feature pipeline.
The securesense.device.v1 stream reports the security state, subtype,
confidence values, temporal vote, safeguard state, telemetry and latency. The
console treats these messages as device evidence rather than recomputing the
final decision.