Adds an `optimize` module that replaces the hand-picked shield config with a
derived, robustness-verified optimum, and hardens the experiment so the
collapse is proven to be signal-level, not classifier-level.
Model changes:
- throughput.rs: add a feedback-airtime term (cost rises with feedback bits)
alongside the falling quantization residual, giving a genuine interior
throughput optimum in feedback resolution.
- attacker.rs: add a selectable distance metric (Euclidean + Cosine) so the
optimizer can require the collapse to hold under multiple classifiers.
- experiment.rs: thread the attacker metric through; build the channel once.
optimize.rs:
- optimal_feedback_bits / spec_optimal_feedback_bits: throughput-best resolution
(3 bits unconstrained, matching DySPAN-2026; 5 bits within the 802.11 {5,7,9}
set).
- min_givens_passes: smallest mixing budget that collapses re-ID robustly across
both metrics AND N in {16,32}.
- pareto_frontier and hyper_optimize.
Findings and adopted defaults:
- Proven-minimum robust passes = 48; the hand-picked 112 was 2.3x over-
provisioned. Rotation mixing is keyed (never signaled), so extra passes are
throughput-free -> ship 96 (2x margin).
- Feedback resolution 5 bits (spec-optimal), down from 7.
- ShieldConfig::default() now equals hyper_optimize()'s output; a test guards
against drift.
Net vs. the original: strictly better on BOTH privacy and throughput.
Reference (SYNTHETIC/L0, N=16): re-ID 100% shield-off -> 4.7% shield-on
(chance 6.25%, below chance), throughput 97.6%, energy ratio 1.000000. 35 tests
+ doctest pass; clippy -D warnings clean; builds for wasm32.
Docs: new docs/research/privacy-shield/08-optimization.md; updated bundle
README/03/05/07 and ADR-288 with the derived operating point.
Co-Authored-By: claude-flow <ruv@ruv.net>
Claude-Session: https://claude.ai/code/session_01WEXNqzs7UsfNFBcP5yW21p
5.0 KiB
05 — Experiment Protocol: Attacker vs. Protector
This is the "start today" deliverable from the brief: make one RuView node the
attacker and one the protector, and measure whether protection drives identity
recognition toward chance while keeping throughput above 95%. It is realized as
a deterministic, reproducible experiment in
v2/crates/wifi-densepose-privshield.
Because it runs on SYNTHETIC data (no radio is touched), its numbers describe
the model, not real hardware — reproduced by cargo test, and to be
re-established on silicon with a captured log before any deployment claim.
1. Setup
- Protector node. Emits beamforming feedback shaped by the VEIL controls (keyed per-session fine-subspace rotation + configured feedback resolution and sounding overhead). Models a legitimate AP/STA protecting a room.
- Attacker node. A passive sniffer that enrolls a template per candidate from captured reports, then classifies fresh captures (nearest-centroid) — the BFId-class re-identification threat.
- Scene.
SceneConfigdefault: 64-dim report, 8 comm dims, 16 candidate identities (chance = 1/16 = 6.25%), per-identity stable fine-block signature- per-session environmental nuisance.
Two runs of the attacker are compared: shield off (the attacker sees raw reports) and shield on (every captured report is VEIL-protected). The same attacker faces both.
2. Metrics and acceptance bar
| Metric | Definition | Bar |
|---|---|---|
| Re-ID accuracy, shield off | Top-1 identity accuracy on unprotected traffic | Must be well above chance (threat is real) — bar ≥ 0.5 |
| Re-ID accuracy, shield on | Top-1 identity accuracy on protected traffic | Must fall into the chance band 1/N · 2 + 0.03 |
| Throughput ratio | Protected link capacity ÷ baseline capacity | ≥ 0.95 |
| Compliance | Emission energy ratio ≈ 1 and non-interfering | is_compliant == true |
Overall passed() requires all four.
3. Results (SYNTHETIC, hyper-optimized default configuration)
Reproduce with cargo test -p wifi-densepose-privshield (all 35 tests + doctest
pass). The default shield config is the optimize module's output — 96 Givens
passes at 5-bit feedback resolution (see
08-optimization.md). Salient values from the reference run:
| Metric | Value |
|---|---|
| Candidate identities | 16 |
| Chance level | 6.25% |
| Chance band (acceptance) | ≤ 15.5% |
| Re-ID accuracy, shield OFF | 100.0% |
| Re-ID accuracy, shield ON | 4.7% |
| Throughput ratio | 97.60% |
| Emission energy ratio | 1.000000 |
| Overall verdict | PASS |
Reading the result: the attacker is a perfect re-identifier without protection (the synthetic signatures are cleanly separable), and VEIL drives it to the chance floor (4.7% sits just below the ideal 6.25%, i.e. no better than guessing) — while the modeled link keeps 97.6% of its throughput and the emission conserves energy exactly (compliant, not jamming). The same collapse holds under a Cosine-metric attacker and at N=32, confirming it is a property of the signal, not the classifier.
4. Determinism and the witness
The experiment is byte-reproducible: no OS entropy, no wall-clock, no threads.
proof::Proof folds the salient outputs (quantized to avoid last-bit f32
round-off) into an FNV-1a witness pinned as EXPECTED_WITNESS. Any drift in the
PRNG stream, rotation schedule, throughput formula, or scene geometry changes the
witness and fails witness_matches_pinned. This is the same
deterministic-proof discipline as nvsim and the Python verify.py.
5. Sensitivity and what to vary next
ExperimentConfig exposes the levers for a fuller study:
scene.identities— larger N lowers the chance floor; confirm collapse holds as candidates grow.scene.env_sigma/beam_amplitude— nuisance and comm energy; stress the separability assumption.shield.feedback_bits— trace the privacy–throughput curve (thethroughputtests already show coarse resolution costs more).shield.givens_passes— mixing strength; fewer passes should degrade the collapse gracefully.- Stronger attacker — swap in a learned classifier to confirm the collapse is signal-level, not classifier-level (the argument says it must be, but a hardware study should verify).
6. Path to a real two-node measurement
The synthetic experiment is the design proof. The hardware path (per CLAUDE.md, requires a captured log to claim MEASURED):
- Two ESP32-S3/C6 or Nexmon-capable nodes: one runs Wi-BFI capture (attacker), one runs a VEIL-shaped feedback profile (protector).
- Enroll and test the same BFId-style classifier on captured BFI, shield off vs. on; log throughput via iperf across the legitimate link.
- Success = the same shape as §3 on real captures, with the boot/runtime log as the witness. Until then, all defense numbers remain SYNTHETIC.