Fusion raises confidence but widens privacy exposure
Aliases: respiration-posture fusion · bodily state dossier · contextual privacy
What it is
Combining respiratory rhythm with posture or activity makes it easier to tell “breathlessness from walking” from “tense breathing while seated.” Confidence rises, and a bodily state dossier is assembled at the same time: when the person lay down, whether they were speaking, whether breathing was disordered. What fusion buys on this channel is not abstract multi-sensor theory; it is inferability of sleep, anxiety, and situation.
Why it happens
Respiration alone is ambiguous: a faster rate may be metabolism, speech, or affect. Posture alone is ambiguous: sitting may be work or staring. Jointly, the conditional probabilities sharpen—walking with faster breath looks metabolic; lying with shallow fast breath looks like sleep or a startle. Depth cameras, radar, and a chest belt, once on together, can also rebuild a timeline of in-room activity. Privacy widens through attribute inference: the user thinks the system is asking “are they at the desk,” while the model can separate conversation, exercise, and night waking. More kinds of data mean that leaking a log hurts as “bodily schedule and possible health state,” not as “step count.”
Studying it
Privacy work often uses an inference attack: given fused features or labels, a second model predicts unauthorized attributes (talking to someone, insomnia). Factors: which sources are fused (posture only, posture plus respiration, plus indoor presence). Outcomes: primary-task accuracy and attack accuracy on sensitive attributes. Interviews should ask what users think is being measured, against the actually inferable set. Reporting only primary F1 hides the privacy cost. Short laboratory wear also understates secondary use of a long-running log.
Where it stops holding
Local, short-lived, non-persisted fusion (suppress a false trigger while walking, do not upload features) exposes less than a cloud archive. In clinical monitoring, high confidence may be explicitly requested, and the informed scope differs. A single noisy source is not automatically safer: repeated false alarms tempt a product to turn more sensors on. In a child’s room or a shared home, posture plus respiration can expose other people in the room; the subject of consent is no longer only the wearer.
Applying it
- List inferable states for “posture × respiration” separately, and allow posture-only with respiration off.
- Default to using fused features to suppress false alarms on the device; do not upload the time series. If the cloud is required, aggregate by day rather than storing each respiratory cycle.
- Show which sources currently participate in fusion, rather than a single “context intelligence” switch.
- Verify by taking an exported product log and checking whether an unauthorized reader can recover sleep bouts or talking bouts; every recoverable item should appear in the informed notice.