Side Eye

A wearable that detects smart glasses (e.g. Meta Ray-Bans) by listening for their Bluetooth Low Energy advertisements, and buzzes when one stays close for too long. Built at HackTheBurgh.

How it works

• A Raspberry Pi 4 scans BLE. For testing, an Android phone running nRF Connect advertises a stand-in packet (company ID 0x058E, data CA01).
• Each second of signal becomes 4 features over sliding windows:
  • mean_rssi: smoothed RSSI, last 5 s
  • var_rssi: raw RSSI variance, last 5 s
  • slope: RSSI trend, last 5 s
  • dwell_s: seconds on the near side of a calibrated cutoff, last 10 s close_streak_s and packet_rate_hz are also logged.
• A 3-second rule drives the vibration motor. A small random forest (20 trees) gives an ML "linger" score.
• Labels: in a linger walk, seconds after the first 3 s of the stop are 1; everything else is 0.

Scope

Side Eye detects smart glasses that are switched on and stay close to you. Bluetooth can't show whether they're recording, or where they're pointed. Design case: glasses worn on a face. Details and walk protocol: docs/DECISIONS.md.

Evaluation

24 real walks: 8 linger, 8 passer, 4 far, 2 bag, 2 blocked. Evaluated on held-out whole walks: 4 folds, each holding out 2 linger + 4 other walks.

Because 24 walks is very little data, we also simulate thousands of walks from public COVID contact-tracing Bluetooth data and compare four methods on the same folds:

Method	Trained on
Rule	3-second rule, no training
Model 1	real walks only
Model 2	simulated walks only
Model 3	simulated + real, simulated rows down-weighted

Team

Person	Part	Folder
Emma Burgess	Pi code: scanner, features, rule, model training, service, tests	pi/
Rachel Ranjith	Public-data augmentation (MITRE + NIST), model comparison	nist/
Lili Berenguer Jin	Pi code (with Emma Burgess)	pi/

Layout

pi/        Raspberry Pi code: scanner, features, rule, training, service, tests
nist/      public-data simulation and model comparison
data/      real walks: data/runs_session1/, data/runs_session2/; public datasets in data/external/ (ignored)
scripts/   setup helpers
gcp/       retraining on Google Cloud: Tiger Data labels + walks -> shape_model.joblib -> Pi (gcp/README.md)
docs/      decisions and scope (docs/DECISIONS.md)

Setup

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

Running each part

Pi (pi/), from the repo root (full details in pi/README.md):

python3 pi/canary.py calibrate            # writes settings.json
python3 pi/canary.py record linger        # one walk -> data/runs/
python3 pi/train_model.py                 # trains model.joblib
python3 pi/canary.py live                 # run the detector

Public-data augmentation (nist/), from the repo root:

scripts/get_mitre.sh                       # clone MITRE, build data/external/mitre_rssi.csv
python3 nist/simulate_walks.py data/external/mitre_rssi.csv settings.json nist/sim --n 300 --rate <Pi packets/s>
python3 scripts/run_results.py --runs data/runs_session1      # all tables, tests and the T25 decision

Details and TODOs in nist/README.md.

Data sources

• MITRE Range-Angle (Structured): github.com/mkrangle/MITRE-Range-Angle-Structured. Only recordings with an Android transmitter (partner_tester in Index.xlsx) are used, since those used nRF Connect like our test glasses. RSSI 127 is an invalid sentinel and is dropped.
• NIST BLE RSSI dataset: data.nist.gov/od/ds/mds2-2827/BLE_RSSI_Data.zip (person-in-between and wall obstruction). Not integrated yet.