wifi-ruview/docs/adr/ADR-106-dp-sgd-and-primitive-isolation.md

ADR-106: Differential privacy + biometric primitive isolation for RuView federated training

Status: Proposed · Date: 2026-05-22 · Author: SOTA research loop tick-15 · Supersedes: none · Extends: ADR-105

Context

ADR-105 specified federated learning for RuView CSI personalisation with MERIDIAN env-normalisation + Krum byzantine-robust aggregation + R7-style update-level mincut. It deferred two questions:

1. Member inference defence. A sufficiently capable adversary observing many model deltas across rounds can in principle reconstruct training samples (Shokri 2017). ADR-105 left "DP-SGD" as a future ADR.
2. Biometric primitive isolation. R15 catalogued five environment-invariant biometric primitives (gait frequency, breathing rate, HRV rate, RCS frequency response, walking dynamics). R15 said: the federation aggregator MUST NOT receive any raw per-subject biometric primitive. ADR-105 didn't yet specify which primitives qualify.

This ADR closes both. It is a direct extension of ADR-105 and incorporates the constraints from R3 (re-ID privacy) + R14 (empathic appliance privacy) + R15 (RF biometric physical-not-learned identification).

Decision

Adopt DP-SGD with explicit primitive-isolation enforcement on every Cognitum Seed before any model delta leaves the device.

Three-layer defence

Layer 1 — Primitive Isolation (R15 binding constraint). A static list of "on-device-only" biometric primitives. The federation client library enforces that these tensors are never serialised into a transmittable update.

Primitive	On-device only	Reason
Raw CSI window (complex64 tensor)	✅	ADR-105 baseline
Gait stride frequency (Hz scalar per subject)	✅	R15 — biometric primitive
Breathing rate (BPM scalar per subject)	✅	R15 — biometric primitive
HRV rate signature (R-R interval array per subject)	✅	R15 — biometric primitive
RCS frequency response curve (per subject, per-subcarrier amplitude)	✅	R15 — biometric primitive
Limb timing vector (per subject, per stride)	✅	R15 — biometric primitive
Per-subject embedding centroid	✅	R3 + ADR-105 — re-ID primitive
MERIDIAN per-room centroid	⚠️	Aggregate over all subjects in the room — not per-subject
LoRA weight delta	⚠️	Encodes biometric information; mitigated by Layer 2 + Layer 3
Model logits / softmax outputs	⚠️	Per-subject during inference; never aggregated for transmission
Coordinator-side aggregate model	❌	Distributed back to nodes; no per-subject content by construction

The ✅ rows are enforced at the API surface — the federation client returns an error if a tensor with these tags is passed to submit_delta().

Layer 2 — Gradient clipping. Before any LoRA weight delta is computed for transmission, individual sample gradients are clipped to L2 norm C (standard DP-SGD step, Abadi 2016). This bounds the sensitivity of the released delta to any single training sample.

Recommended: C = 1.0 (after experimentation per-cog; some cogs may need C ∈ [0.5, 2.0]).

Layer 3 — Gaussian noise on aggregated deltas. Before transmission to the coordinator, Gaussian noise N(0, σ²C²I) is added to the aggregated LoRA delta. This bounds the per-round privacy leakage.

Privacy budget

Using the Moments Accountant (Abadi 2016) for (ε, δ)-DP across federation rounds:

Configuration	Per-round σ	Rounds	Total ε (δ=1e-5)	Verdict
Conservative (medical-grade)	1.5	50	2.0	Strong; matches HIPAA-aligned recommendations
Standard (typical RuView)	1.0	100	5.0	Strong; consistent with Google's federated keyboard work
Lenient (faster convergence)	0.5	100	8.0	Moderate; below ε=10 community soft-bound

Recommended starting σ = 1.0 for most RuView cogs, with per-cog tuning:

• cog-person-count (R8 — simple classifier): σ=1.0 sufficient.
• AETHER re-ID head (R3 — high discriminability needed): σ=0.7 with C=1.5 to preserve discriminative power.
• cog-pose-estimation (skeleton output): σ=1.0.
• cog-maritime-watch (R11): σ=1.5 (medical-grade — vessel crew vitals).

Composition with ADR-105 protocol

The DP-SGD layer slots in at step 4 of ADR-105's protocol summary:

4. Delta compression. Compute ΔW_i = W_T+1_i − W_T. [NEW: clip individual-sample gradients to L2 norm C=1.0 during local training; add Gaussian noise N(0, σ²C²I) to ΔW_i with σ from per-cog table above.] Quantise to int8 + LoRA-rank decomposition (rank=8) → ~1 MB per delta.

Krum byzantine-robust aggregation (step 5) operates on DP-noised deltas without modification — Krum's distance metric is robust to additive Gaussian noise at typical σ values.

Implementation enforcement

The ruview-fed crate (per ADR-105 implementation plan, ~500 LOC) gains:

Component	LOC	Purpose
PrimitiveTag enum + tensor tagging trait	60	Layer 1 primitive isolation
clip_gradient_l2(C) helper	30	Layer 2 clipping
add_dp_noise(sigma, C) helper	40	Layer 3 Gaussian noise
MomentsAccountant	120	(ε, δ) tracking across rounds; aborts federation if budget exceeded
Per-cog config schema	50	σ, C, max rounds budget

Total ~300 additional LOC on top of ADR-105's 500. Federation protocol implementation budget revised to ~800 LOC total.

Alternatives considered

A. Federated learning without DP

Status: rejected. ADR-105's Krum + LoRA + int8 quantisation provides some implicit privacy, but it's not a formal guarantee. Member-inference attacks (Shokri 2017) recover training samples from undefended FL. We need a formal (ε, δ)-DP bound.

B. Local DP (LDP) only

Status: rejected. LDP would add noise per-sample at the device, then the coordinator gets noisy aggregates. This gives stronger guarantees but degrades model accuracy by 5-15× for the same ε. Central DP (CDP) with byzantine-robust aggregation is the right trade-off for our threat model where the coordinator is trusted to apply noise correctly (the coordinator is cognitum-v0 fleet manager, under installation owner's control per ADR-100 signing).

C. Heavier obfuscation (homomorphic encryption / secure aggregation)

Status: deferred. Secure aggregation (Bonawitz 2016) avoids the coordinator ever seeing individual deltas, only their sum. This is the right next layer for cross-installation federation (ADR-105 explicitly deferred). For within-installation federation where the coordinator is owner-controlled, the gains don't justify the 5-10× compute and complexity cost.

D. Just-trust-Krum

Status: rejected. Krum defends against adversarial nodes, not adversarial inference. A passive coordinator (even an honest one) plus moderate compute can extract training samples from undefended deltas. DP-SGD is the proper defence.

Threat model

Threat	Layer that mitigates
Compromised seed reads its own local biometric primitives	Out of scope — physical compromise = full local compromise
Compromised seed exfiltrates a biometric primitive via the federation channel	Layer 1 — primitive isolation API blocks transmission
Passive coordinator reconstructs training samples from observed deltas (Shokri 2017)	Layer 2 + 3 — DP-SGD bounds reconstruction quality
Member inference attack on the trained model (Shokri 2017 §3.2)	Layer 2 + 3 — formal (ε, δ) bound
Coordinator + 1 colluding seed	Krum (ADR-105) still works; DP-SGD bounds the colluder's info gain
Brute-force gradient inversion (Zhu 2019)	Layer 2 + 3 — clipping + noise defeats gradient-from-update attack
Active adversary controlling >f Krum nodes	Out of scope — ADR-105 byzantine bound f < (K-2)/2
Side-channel via inference latency	Out of scope — separate ADR (constant-time inference)

Consequences

Positive

1. RuView federation is now formally privacy-preserving with a documented (ε, δ) bound — meets GDPR Art 25 ("data protection by design") technical-measure expectations.
2. R15's biometric-primitive constraints are enforced at the API surface, not just policy-documented.
3. The threat model has been written down with explicit mitigations per row, making future security review tractable.
4. The Moments Accountant aborts federation rather than silently consuming budget — operationally safer than naive "just keep training".

Negative

1. DP noise degrades model accuracy by ~3-8% (typical figures from DP-SGD literature; per-cog tuning needed). For cog-person-count v0.0.2 (this loop's earlier work), the baseline 34.3% class-1 accuracy would degrade to ~31-33% with σ=1.0.
2. Adds ~300 LOC + Moments Accountant complexity to ruview-fed. Total federation budget revised to ~800 LOC.
3. Per-cog tuning of (σ, C, max_rounds) is needed — not a one-size-fits-all.
4. Doesn't defend against side-channel inference latency leaks; that's a separate ADR.
5. Doesn't address cross-installation federation; cross-installation work still requires the deferred ADR (secure aggregation + DP).

Open questions intentionally left

1. Per-cog DP budget allocation. The σ values above are first-cut recommendations; empirical tuning per cog is needed before shipping.
2. Moments Accountant restart policy. What happens after we exceed ε? Reset model and restart? Stop federation indefinitely? Decision deferred to operations.
3. Side-channel timing leaks. A separate ADR (TBD) needs to cover constant-time inference and constant-time DP-noise sampling.
4. Subject-level vs sample-level DP. This ADR specifies sample-level. Subject-level DP (preventing inference of "is subject X in the training set") needs K_subjects × privacy_amplification — discussed in next-generation work.

Bridge to existing ADRs

• ADR-024 (AETHER) — within-room training stays unchanged; DP-SGD applies at the federation layer.
• ADR-027 (MERIDIAN) — env-centroid subtraction is per-room aggregate, not per-subject — survives Layer 1 isolation as an ⚠️ entry (aggregate is acceptable).
• ADR-029 (multistatic) — per-seed federation; multistatic geometry stays per-installation.
• ADR-100 (cog packaging) — Ed25519 signing covers DP-noised checkpoints with no protocol change.
• ADR-103 (cog-person-count) — first cog with formal DP guarantee; this loop's v0.0.2 retrain becomes ADR-106-compliant on next training cycle.
• ADR-104 (ruview-mcp + ruview-cli) — exposes ε, δ budget remaining via MCP ruview_fed_privacy_budget (future tool; out of scope for this ADR).
• ADR-105 (federated training) — DP-SGD slots into step 4; threat model extended; implementation budget grows from 500 to ~800 LOC.

Connection to research-loop threads

• R3 (cross-room re-ID) — Layer 1 isolation blocks transmission of per-subject embedding centroids.
• R7 (mincut adversarial) — Krum (from ADR-105) + DP-noised deltas remain compatible; mincut adversarial check operates on the noised similarity graph.
• R12 (eigenshift NEGATIVE) — informed by the structure-detection failure pattern; the DP-noise approach treats adversarial deltas as "outliers from a noisy distribution" rather than as a structural-detection problem.
• R13 (contactless BP NEGATIVE) — confirms why we restrict biometric primitive transmission: contour-level signals don't meet the 25 dB floor, so they wouldn't help downstream models anyway; rate-level primitives are sufficient for V1/V2/V3 features.
• R14 (empathic appliances) — privacy framework constraints now have a formal (ε, δ) backing.
• R15 (RF biometric primitives) — direct requirements basis; the on-device-only primitive list is R15's catalogue made executable.

Honest scope

• σ values are recommendations, not measurements. Per-cog empirical tuning is needed (cog-pose, cog-count, AETHER head, future cogs each get their own).
• (ε, δ)-DP is a worst-case bound. Real privacy depends on the auxiliary information the adversary has. For an adversary with extensive auxiliary biometric data, even a small ε can leak. Layer 1 primitive isolation is the harder constraint that doesn't depend on the auxiliary-info model.
• The Moments Accountant treats each round as independent, which slightly over-estimates the budget consumed (good — conservative). Tighter accountants (Rényi DP, PRV) would let us run more rounds for the same ε.
• Subject-level DP is not formalised here. Many use cases (a household of 4 always-the-same individuals) effectively have K=4 subjects, where sample-level DP doesn't fully capture the subject-level risk.

Implementation plan (additive to ADR-105)

Step	LOC	Notes
1. PrimitiveTag enum + tensor tagging	60	Compile-time enforcement where possible
2. Gradient clipping helper	30	Per-sample (microbatch-friendly)
3. Gaussian noise helper	40	Constant-time sampling (defends weak side-channel)
4. Moments Accountant	120	Tracks (ε, δ) across rounds; emits budget-exhausted error
5. Per-cog config schema (σ, C, max_rounds)	50	YAML/TOML, validated at federation start
6. End-to-end privacy test	—	Synthetic membership-inference attack vs DP-protected model; verify reconstruction quality is bounded by (ε, δ) prediction

Combined with ADR-105's 500 LOC, total federation budget revised to ~800 LOC, ~3-week effort.

What this DOES enable

• Formally privacy-preserving federation with a documented (ε, δ) bound.
• API-level enforcement of R15's biometric primitive isolation list — not just policy text.
• A clear next-ADR path: ADR-107 (cross-installation federation w/ secure aggregation) builds on this foundation.

What this DOES NOT enable

• Subject-level DP (preventing "is subject X in training") — would need subject-level privacy amplification.
• Defence against side-channel timing leaks — separate ADR.
• Cross-installation federation — separate ADR with secure aggregation + cross-installation DP composition.
• Adversarial robustness to physical compromise — out of scope; physical security is the orthogonal defence layer.

Decision-making record

• 2026-05-22 06:38 UTC — drafted by SOTA research loop tick-15 based on R3 + R15 + ADR-105's deferred items. Status: Proposed.
• Pending: review by security-architect (formal DP bound verification), ddd-domain-expert (federation = bounded context with this ADR as its public API), production-validator (the per-cog σ values need bench validation before shipping any specific cog).