R3 — Cross-room CSI re-identification: AETHER + MERIDIAN synthesis

Status: simulation + ADR-024/027 synthesis + privacy framing · 2026-05-22

The question

AETHER (ADR-024) gives us contrastive CSI embeddings that achieve ~95% within-room 1-shot re-identification on MM-Fi. Can the same embeddings identify the same person across a different room?

This question has two answers — a technical one and an ethical one. R3 takes both seriously.

Decomposition

A CSI embedding from any frame is approximately:

embedding = person_signature + environment_signature + noise

The environment signature includes multipath geometry, AP placement, furniture, walls. It is constant per (room, antenna placement), and changes by O(1) between rooms — empirically larger than the per-person signature variation. This is exactly the structure that ADR-027 (MERIDIAN) targets.

examples/research-sota/r3_crossroom_reid.py simulates the problem with physics-realistic parameters: 10 subjects, 3 rooms, 128-dim embeddings, person-signature scale 0.35, environment scale 1.5 (env ≈ 4.7× person), noise 0.3.

Results

Three observations:

1. Cosine K-NN partially mitigates the environment-shift problem (70% >> 10% chance) because magnitude normalisation removes the additive env component as a direction. The remaining 30 pp gap comes from how the env shift rotates the cluster in the high-dim space.
2. Explicit MERIDIAN-style env subtraction (per-room centroid removal) closes the remaining gap. The simulation suggests even 70%-effective subtraction (realistic for finite labelled examples) is enough.
3. The within-room baseline is what an attacker has, not what the system needs. The same primitive that gives the user "let RuView greet you by name in this room" also gives an attacker "this person walked through 5 different rooms and we tracked them."

Why the env-removal approach works

MERIDIAN's core idea (ADR-027) is to estimate environment_signature from labelled samples in the new room and subtract it. The estimator works because:

• All people contribute equally to the per-room mean (assuming reasonably balanced training data)
• The person signatures are zero-mean across the population (an embedding is meaningful only relative to others)
• Therefore mean(embeddings in room R) ≈ environment_signature[R]

Subtracting the per-room centroid gives embedding_clean ≈ person_signature + noise, which is the room-invariant signature.

Trade-off: MERIDIAN needs labelled (or at least clustered) examples in the new room to estimate its centroid. Pure zero-shot transfer to an unobserved room is much harder — without any anchor, you can't distinguish "person A in new room" from "person B in old room" robustly.

Physics gives us another lever

R6's Fresnel forward model tells us where the env_sig lives in the embedding: it's the contribution from the multipath / reflector geometry. A 5 m bedroom has 4-6 dominant reflector positions; the env_sig is a function of those.

If we could predict the env_sig from the forward model + a room geometry (R6's A matrix + a coarse map of the room), we wouldn't need labelled examples. This is the next-tier sophistication: physics-informed domain invariance rather than statistically estimated.

This isn't built. It's the right next step in the AETHER + MERIDIAN line.

Privacy framing (the ethical answer)

The same primitive that enables "RuView greets you by name in your bedroom" enables a building-level adversary to track every individual's movement through every WiFi-CSI-sensing surface. This is a stronger surveillance primitive than face recognition because:

• WiFi penetrates walls (no line-of-sight needed)
• Re-ID works without subject cooperation (no "look at the camera")
• The signal is invisible (no light, no observable signal)
• The biometric is the body's RF signature, not a removable accessory

The R14 ethical framework (opt-in by default, data stays on-device, override is one tap) applies, but with additional constraints specific to re-ID:

1. No cross-installation linkage. Per-installation embedding spaces only. Two RuView installs in two different buildings must NOT share embedding spaces.
2. Embedding storage requires explicit opt-in. Storing person embeddings persists biometrics; many regulatory regimes treat this as biometric data with stronger consent requirements (GDPR Art 9, BIPA).
3. Forgetting must be cryptographically verifiable. When a user requests deletion, the embedding must be cryptographically destroyed, not just unlabelled. Storing "unlabelled embeddings" still enables future linkage.
4. No re-ID across legal entities. Building A and Building B owned by different entities must NOT exchange embeddings. The data-flow boundaries should be hard-walled.

These constraints make some use cases impossible (e.g. "automatic global biometric ID" — yes, that's the point) and some clearly aligned with the user (e.g. "remember which family member is in which room").

What this enables

1. Per-installation personalisation — empathic appliances (R14) get per-person calibration after MERIDIAN-style env subtraction.
2. Anomaly detection — "someone walked into this room who isn't in the household's embedding set" → home-security primitive without face recognition.
3. Pose-data-association — multi-person pose tracking in the same room can use the embedding to maintain consistent identity through occlusion.

What this DOES NOT enable (correctly, by design)

1. Cross-building tracking
2. Re-ID across legal entities
3. Long-term unlabelled biometric storage
4. Zero-shot transfer to unobserved rooms (without physics-informed extension)

Honest scope

• The simulation uses additive person + env + noise decomposition. Real CSI has multiplicative environment effects in the multipath domain — env modulates person signature amplitude in subcarrier-specific ways. A more realistic forward model would multiply the per-subcarrier slot transfer function with the person signature, which makes env-removal harder (not just subtraction).
• The 70% cross-room raw cosine K-NN number depends heavily on env / person scale ratio. With a 10× larger env (e.g. crossing from a bedroom to a kitchen with very different multipath), the raw cosine K-NN drops further. With a 2× smaller env (very similar rooms), it barely drops. The MERIDIAN closing of the gap appears robust.
• We did not simulate adversarial scenarios where an attacker actively manipulates the env signal to break tracking. R7's mincut would have to weigh in on this.

Connection back

• R5 (saliency) — within-room saliency profiles include both the person- and environment-saliency. Cross-room transfer would need to find the person-only saliency, which is a research problem AETHER (ADR-024) partially addresses through contrastive learning.
• R6 (Fresnel) — the missing piece: physics-informed env_sig prediction from a room model. Not yet built.
• R7 (mincut adversarial) — cross-room re-ID is the highest-risk surface for adversarial spoofing. If the system can be fooled into thinking "person B is in room A", that's a security incident; multi-link consistency from R7 is the defence.
• R9 (RSSI K-NN) — already showed that even RSSI alone preserves a weak locality signature within room; the cross-room transfer for RSSI is worse than for full CSI, but the env / person decomposition still applies.
• R14 (empathic appliances) — re-ID enables per-occupant V1 lighting / V2 HVAC / V3 attention-respecting. The privacy constraints from R14 + the four cross-installation constraints from R3 together are the binding spec.

Next ticks (R3 follow-ups)

• Physics-informed env_sig prediction from R6's forward operator + a coarse room map → zero-shot cross-room transfer.
• Multi-occupant re-ID under occlusion: two people in the same room, intermittent visibility of each; can a Kalman + AETHER pipeline maintain identity continuously?
• Cryptographic forgetting protocol: how do you prove an embedding has been deleted to a regulator who can't see your hard drive? (Out of scope for this loop, but a real research question.)