Files

27 KiB
Raw Permalink Blame History

title, status, date, scope
title status date scope
Synthesis: the paper we can defend research-synthesis 2026-11-14 PulseCast / Future City Hackathon 2026; extended paper, not the 4–5 page hackathon report

Synthesis: the paper we can defend

0. One-sentence thesis

The PulseCast build proves a decision loop, but its sensing layer is not a people counter: each “15-minute” value is a ~10-second count of randomized BLE addresses, truncated to 16 bits and multiplied by a fixed 0.7 factor; the defensible paper is therefore a code-level audit of that claim plus a privacy-first, decision-grade architecture for the next version — not a claim that eight sensors measure a city.

1. What we now know

1.1 The measurement chain (code-level, high confidence)

The deployed Code for Heilbronn lorax firmware (sensor-pax) does this:

wake from deep sleep
  → BLE passive scan (window 80 ms, interval 80 ms)
  → count distinct locally-administered BLE addresses in a ~10 s window
  → truncate the last 2 address bytes to a 16-bit ID
  → deduplicate within that scan only
  → multiply the integer count by factor (default 0.7)
  → send one LoRaWAN payload: sensor_id, 0, value, 1, battery
  → deep sleep 900 s

Key code facts:

  • configuration.blecounter = 1, configuration.wificounter = 0 — BLE only.
  • configuration.blescantime = 30, but the vendored start_BLE_scan() does not use blescantime; the main loop stops on the first 10 s report callback. Effective scan ≈ 10 s, not 30 s.
  • libpax_counter_init(log, &count, 10, 1) — 10 s report timer, cumulative count within the wake.
  • mac_add() accepts only addresses with the locally administered bit set; it ignores universal/static addresses.
  • id = (uint16_t *)(paddr + 4) — the “hash” is a 16-bit truncation of the last two bytes, not a cryptographic hash.
  • paxCount * cfg.factor with factor default 0.7; the value is formatted with four decimals.
  • sleep_time_sec default 900 s.
  • RSSI threshold −80 dBm; BLE advertising queue size 30; queue drops are logged, not counted in the payload.
  • The backend stores the parsed value directly; the CSV is a raw export, not a 15-minute mean.

Formula: published_value ≈ 0.7 × U, where U is the number of distinct 16-bit IDs from randomized BLE addresses seen in a ~10 s scan. The sampling interval is ~15 min. The duty cycle is ~1.1% (10 s / ~910 s), not 3.3%.

Sampling consequence: a passer-by with a 30 s dwell has roughly a 4.4% chance of overlapping a scan; a one-hour lingerer can be counted up to four times. The metric conflates flow (people passing) with dwell (people staying), and the dedup bitmap resets every wake, so neither is measured cleanly.

Randomization nuance: within a 10 s scan, BLE address rotation is rare (RPA rotation is typically 900 s, measured 15–45 min). MAC randomization therefore does not mainly inflate the within-scan count; it destroys cross-scan/cross-sensor deduplication and re-identification. The locally-administered-bit filter is also a heuristic: it excludes public/static-address devices and includes non-phone random advertisers (TVs, beacons, earbuds, car head-units).

1.2 Forensic confirmation (data-level, high confidence)

The 8-sensor production window (2026-06-12 → 2026-11-07) contains 83,156 density values. Every one is an exact multiple of 0.7. The reconstructed integer U has median 12, p95 71, max 418, and 2.7% zeros (from the reliability note; verify independently before submission).

The legacy density-01 device 863f75b0 (2025-10-16 → 2026-03-12) is integer-valued; its replacement f8c19730 (2026-04-16 → 2026-11-07) is 0.7-scaled. This is a firmware/calibration break, not a footfall change. Any model trained across the break fits the instrument, not the city.

Additional data-quality findings from the empirical analysis (to verify and cite):

Finding Value Implication
Panel slots with all 8 sensors reporting 28.2% of 13,728 15-min slots; modal count is 7 locations (4,117 slots) The “8-sensor panel” is sparse; a random heatmap often has fewer than 8 points
All-8 coverage when relaxed to 30 / 60 / 120 min 63.4% / 69.7% / 71.0% Temporal aggregation hides the sparsity; it does not create measurements
Worst monthly coverage 23.3% at one location; August 2026 averages 54.2% Coverage is non-stationary; six of eight locations degrade together
All-sensor zero / <1 / <5 share 5.2% / 10.4% / 39.6% Zero-inflation and low-count dominance; log/negative-binomial models needed
Cross-sensor Pearson correlation mean 0.58, range 0.13–0.90 Sensors are correlated but not redundant; the field is not 8 independent measurements
Cross-sensor CV at the same timestamp median 0.66, p90 1.05 Large spatial variability; a single network-wide index hides local differences
Hourly shape correlation across sensors mean r = 0.79 Relative daily shape is consistent; absolute levels are not
Peak hour by sensor 16, 19, 20, 21 “Peak at 19:00” is not universal
Weekend/weekday ratio by sensor 0.81–1.24 “Saturday busiest” is not universal
Missingness bias ratio median 0.69, range 0.38–1.49 Gaps are not random; observed counts differ systematically from full-coverage periods
Battery median / share <20% 17% / 51.3% Power is a plausible cause of dropouts; battery health is part of data quality
Longest flatline run 20 readings / 5 h Stuck-sensor or repeated-value risk
Near-duplicate intervals <60 s 113 Duplicate/reboot artifacts; needs dedup rules

These are empirical findings from the delivered CSV; they do not require new sensors. They make the “local pulse map, not city-wide surface” claim concrete.

What the aggregates cannot tell us: MAC-randomization rate, penetration rate, unique-device count, dwell time, repeat visits, detection radius, or false-positive/false-negative rates. The 0.7 grid is a proxy for firmware/vendor processing, not a direct measurement of randomization or penetration. The paper must not infer these from the CSV.

1.3 Reliability failure modes

Failure Mechanism Consequence
Short snapshot ~10 s scan every 15 min Misses peaks between scans; target is a sparse sample, not integrated flow
Randomized-only filter Only locally administered addresses counted Static/public-address devices invisible; non-phone gadgets with random addresses counted
MAC randomization RPA rotation ~15 min; rare within a 10 s scan Does not mainly inflate within-scan counts; it prevents cross-scan/cross-sensor dedup and re-identification
16-bit truncation Last two bytes used as ID Collisions grow with density; undercount at busy times
No cross-scan dedup Bitmap reset every wake Same person counted again in every 15-min bin; dwell not measurable
Device ≠ person Phones, watches, earbuds, cars, beacons Unknown multiplier; one person can be several devices
Penetration unknown Bluetooth on/off, discoverability, demographics Absolute counts not identifiable without calibration
RSSI gate −80 dBm threshold, body shielding, orientation Detection zone is not a stable catchment
Queue drops 30-slot advertising queue Undercount under density; not exposed in data
Gaps LoRaWAN joins, power, maintenance 4–6 day gaps; missing not at random
Factor Default 0.7, undocumented Single scalar cannot absorb penetration × devices/person × scan probability

1.4 Privacy and law

The architecture is better than raw MAC retention: raw addresses stay in RAM, only counts leave the sensor. But:

  • The 16-bit truncation is not a cryptographic hash; it is a dedup key with a small ID space.
  • The operator’s blog says MACs are “hashed and stored for at most one hour”; the public 2026 firmware resets the bitmap every wake and has no one-hour store. The two accounts disagree; report both.
  • MAC addresses are very likely personal data in this context: a device identifier plus sensor location, time, and deduplication can single out a device and its bearer. [CJEU] Breyer, C-582/14, paras 42–46; GDPR Art. 4(1), Recital 26; EDPB Guidelines 02/2026 (draft), para. 25.
  • Hashing or truncating a MAC is pseudonymisation, not anonymisation. The Conseil d'État's JCDecaux decision held that truncation + salt + keyed hashing still left the controller able to identify and link records; the Dutch Enschede Wi-Fi tracking fine was annulled on evidential grounds, not because the law was settled. [Enforcement] CE, 8 Feb 2017, no. 393714; ECLI:NL:RBOVE:2024:594; ECLI:NL:RVS:2026:4403.
  • Passive sniffing under ePrivacy Art. 5(3) / § 25 TDDDG is contested: the better textual reading is that passive reception is not “access” to terminal equipment, but the EDPB's functional approach creates material risk; if § 25 applies, prior consent is required and legitimate interest/signage are not enough. [Law] ePrivacy Art. 5(3); TDDDG § 25; [Guidance] EDPB Guidelines 02/2023, paras 39, 42; DSK OH Digitale Dienste (Nov 2024).
  • A DPIA is almost certainly mandatory for continuous sensing of a publicly accessible inner-city area (Art. 35(3)(c)); the Baden-Württemberg private-sector “Muss-Liste” explicitly covers offline tracking of customer movements and radio-signal location over a substantial period. [Law] GDPR Art. 35(3)(c); [Guidance] LfDI BW Muss-Liste, Nos. 2 and 10.
  • German enforcement has been administrative rather than fine-based for this use case: the ULD Schleswig-Holstein investigated two city-marketing operators using passive probe requests and concluded there was no legal basis; both stopped. The Bavarian authority inspected a fashion chain's Wi-Fi tracking test; the company stopped. [DPA] ULD TB 2019, §5.4.8; BayLDA TB 2015/16, §§3.10, 22.10; HmbBfDI TB 2014/2015, §V.9.
  • “Privacy-preserving” is an architectural claim, not a formal guarantee. A defensible claim needs a threat model, a privacy unit, a k-threshold or DP budget, a retention policy, a DPIA, and a legal basis. If § 25 cannot be excluded, the viable route is a sensor that does not extract MAC addresses at all.

Recommended privacy stack (from the PET note):

  • No raw MACs; edge counting with a rotating, destroyed salt; no cross-sensor or cross-day linkage.
  • Central DP at a trusted city aggregator is the realistic pilot: device-day privacy unit, contribution bounds (e.g. at most four sensor-bin pairs per device per day), fixed daily/weekly coarse releases, and a published budget ledger.
  • Do not use local DP for sparse 15-min bins: its error scales as 1/(ε√n) versus central 1/(εn), so it destroys utility at small-city scale. At ε = 1, a 14-device cell carries ±3 noise; a live 15-min public heatmap is not viable with a meaningful annual budget.
  • k-anonymity is a secondary check, not the primary guarantee: four spatio-temporal points identify ~95% of people in a 1.5M-person mobility dataset; generalizing to k = 20 erases the local footfall signal.
  • Secure aggregation/MPC/HE/TEE complement DP but protect data in computation, not the released statistic; a single city-owned aggregator does not need them in the pilot.
  • Federated learning is not applicable to passive BLE sniffing: the pedestrian's phone does not run our code.
  • Non-DP synthetic data has no formal guarantee; use DP synthesis plus membership-inference/Anonymeter evaluation for public release, not for the live dashboard.
  • Publish the guarantee and the limits: state the privacy unit, ε/δ, contribution bound, release granularity, and what cannot be guaranteed (fine-grained public 15-min heatmaps, individual trajectories, repeat-visit metrics).

1.5 What actually measures footfall better

The alternatives note gives measured ranges. The short version:

Modality Strength Weakness Role
Thermal/IR array 89–93% doorway occupancy; low privacy risk; €150–400/node Narrow passages; degrades >30 °C, >7.5 m, groups Default at store entrances and narrow gateways
mmWave radar 96% presence, 69% count outdoor; low privacy risk; €300–700 Groups, rain, ~8 m range; count error too high without calibration Wide, exposed crossings
Edge AI camera 13.8% day / 29.8% night MAPE outdoors; 2–14% in official review High privacy cost; lighting/occlusion; DPIA 2–4 reference/validation sites
LiDAR >97% disaggregate count; 1.6–4.3% 15-min error €1k–8k; installation; eye safety 1–2 high-value reference sites
Aggregated probe data City-wide movement, Google/HERE/TomTom scale Non-probability sample; penetration/bias unknown; privacy-enhanced, not anonymous Context and city-wide trends, not street-level truth
Manual counts Ground truth; ~11% clicker error Labour; sampling; not continuous Quarterly calibration audits
Sensor fusion Bayesian hierarchical model combines modalities and propagates uncertainty Engineering; identifiability; drift The core of the “right way”

Recommended stack for Heilbronn: 8–12 thermal nodes at narrow entries; 1–2 mmWave/LiDAR nodes at wide/open locations; 2–4 edge cameras as reference/validation; aggregated probe data for city-wide context; twice-yearly manual/video ground truth; transit validators as covariates. No MAC sniffing, no face recognition.

Pilot protocol: 12 weeks at six sites; go/no-go at MAPE ≤ 20%, |bias| ≤ 10%, and 80% interval coverage ≥ 75%. Year-1 cost estimate €81k–161k, recurring €30k–50k [ESTIMATE]; DPIA under Art. 35(3)(c) GDPR and BDSG §4; no raw video or point clouds retained.

1.6 Heatmap and forecast

  • Eight sensors over ~580 m × 280 m give ~142 m equivalent spacing in 2D, or ~72 m along a single street. A 100–150 m support is defensible; a 25–50 m raster is visualization, not inference.
  • Use a negative-binomial hierarchical model with a low-rank spatial field (Matérn GP/SPDE or spatial basis functions), on the walkable network, not isotropic Euclidean smoothing over buildings.
  • Forecast with a global quantile gradient-boosted model (calendar, weather forecast, holidays, events, lags, sensor identity) and conformal calibration. Deep spatiotemporal models and time-series foundation models are benchmarks/ablations at this data scale, not the default.
  • The largest leakage risk is weather: training on observed weather and serving with forecast weather is optimistic. Report both the observed-weather oracle and the forecast-weather number.
  • Evaluate at the decision level (staffing cost/coverage, overflow risk, ad exposure stability), not only MAE/WAPE/R². A point forecast without an interval is not decision-grade.

1.7 Generalization across the OKF_Vault challenges

The applications matrix supports a conditional claim:

Class Challenges What transfers
Direct Staffing, advertising Same sensing/field/forecast; new decision layer
Conditional Waste routing; store location Temporal/decision pattern transfers; waste needs a graph field and geometry; store location needs city-wide data (not supported by 8 sensors)
Sibling Sensor placement (AI TRAQC); LoRaWAN planning (HNVG) Spatial inference and design evaluation transfer; the decision changes the observation operator; no temporal forecast
Non-application Cable-car sentiment; learning guide No physical field, no forecast target, no resource-allocation decision

This is stronger than “one model for six challenges” because it states the boundary conditions.

2. The paper’s contributions

  1. A code-level audit of the deployed PAX counter firmware and backend, with the exact measurement chain (10 s scan, 0.7 factor, 16-bit truncation, randomized-only filter, 1.1% duty cycle).
  2. Forensic data evidence: 100% of 83,156 values are multiples of 0.7; the density-01 calibration break; the heterogeneous gaps and zero-inflation.
  3. A privacy/legal analysis: why “privacy-preserving” is architectural, not formal; the 16-bit truncation is not a hash; the legal basis and DPIA requirements.
  4. A privacy-first reference architecture: edge aggregation, local DP/secure aggregation, k-threshold, no raw trajectories, retention limits; a sensing stack that uses thermal/radar/camera/probe/manual data instead of MAC sniffing as the default.
  5. A sparse-to-heatmap-to-decision pipeline: network-aware Bayesian field, quantile GBM forecast, conformal calibration, decision-level evaluation.
  6. A transferability matrix across the six challenges, with explicit non-applications and falsifiers.
  7. A staged roadmap: what to keep from the hackathon build, what to replace, what to measure before claiming ROI.

3. Proposed structure (extended paper, ~10–12 pages)

Section Pages Content
Abstract 0.25 Problem, audit finding, alternative, transferability, non-claims
1. Introduction 1 Decision gap; the “privacy-preserving” claim; contributions
2. The measurement problem 2.5 Code audit; formula; data forensics; reliability taxonomy; existing PulseCast metrics reframed
3. The privacy problem 1.5 GDPR/ePrivacy/TTDSG; threat model; 16-bit truncation; operator claims vs code; DPIA checklist
4. What “privacy-first” should mean 2 PETs; sensing alternatives; calibration; reference sensing stack
5. From sparse observations to decisions 2 Heatmap; forecast; uncertainty; decision layer; evaluation harness
6. How far it generalizes 1.5 Applications matrix; case studies; falsifiers; where it stops
7. Evaluation agenda 1 Ground truth; calibration; duty cycle; penetration; privacy budget; decision regret; proposed experiments
8. Discussion & limitations 0.75 Single site; no ground truth; opinion bias; legal not advice; hackathon context
9. Related work 1 PAX/privacy; pedestrian forecasting; spatiotemporal models; MLOps; urban dashboards
10. Conclusion 0.5 What to keep; what to build; a checklist
References 1.5 25–35 real, verified sources

For the 4–5 page hackathon report, keep the existing structure and add a 0.5–1 page “what we would do with more resources” subsection. Do not try to fit the extended audit into 5 pages.

4. Claims and non-claims

Supported by evidence:

  • The published value is 0.7 × U, where U is a short-window count of randomized BLE addresses. [code + data]
  • The production 7-day model reaches MAE 3.32, WAPE 27.5%, R² 0.777 on the 28-day holdout; the day-ahead model reaches 3.36/27.8/0.763. Per-sensor MAE ranges from 2.26 (density-01) to 5.83 (Stadtgalerie), so the pooled model is not a homogeneous field. [measured; metrics.json]
  • The 8-sensor network is a local inner-city cluster, not a city-wide sensor. [data]
  • The existing model improves on seasonal naive in a 28-day holdout: MAE 3.36 vs 5.40; WAPE 27.8% vs 44.7%; R² 0.779 vs 0.462. [measured]
  • The sensor counts devices, not people; penetration is unknown. [data + literature]
  • MAC-based sensing has a contested legal basis and is not a formal anonymization guarantee. [law + PET]
  • Thermal/radar/camera/probe/manual modalities have different accuracy, cost, privacy, and coverage trade-offs. [literature]
  • The stack transfers directly to staffing/ads, conditionally to waste, as a sibling to sensor placement/LoRaWAN planning, and not at all to sentiment/learning guide. [applications matrix]

Not supported:

  • “The sensor counts people.”
  • “The value is a 15-minute count or mean.”
  • “The 0.7 factor is a validated penetration correction.”
  • “The 16-bit hash is anonymous.”
  • “Eight sensors support a city-wide heatmap or store-location ranking.”
  • “One model solves all six challenges.”
  • “The proposed architecture has been deployed and evaluated.” (It is a design/research agenda.)
  • “The staffing plan has proven ROI.” (No POS/labor/outcome data.)

5. Figures and tables

Figures

  1. The PAX counter cycle: wake → ~10 s scan → 0.7 × U → LoRaWAN → 900 s sleep. Include the 1.1% duty cycle.
  2. Value quantization and the density-01 calibration break: integer series vs 0.7-scaled series.
  3. Sensor map with coverage/gaps and the 580 m × 280 m cluster.
  4. Privacy-first sensing and data-flow architecture (edge aggregation, DP/secure aggregation, k-threshold, no raw trajectories).
  5. Sparse-to-heatmap-to-decision pipeline (network-aware Bayesian field → quantile forecast → decision layer → evaluation).

Tables

  1. Code audit: what the firmware does vs what the challenge claims.
  2. Data quality: per-sensor coverage, gaps, zero share, reconstructed U, quantization.
  3. Sensing alternatives: accuracy, cost, privacy, deployment constraints, best use.
  4. Model comparison: seasonal naive / no weather / weather; existing PulseCast metrics.
  5. Transferability matrix across the six challenges.
  6. Roadmap: demo / pilot / production; what to build, what to buy, triggers, estimates.

6. Proposed experiments (post-demo, optional)

These are real analyses we can run on the existing CSV; none require new sensors.

  1. Quantization audit: confirm 100% multiples of 0.7 in the 8-sensor window; reconstruct U; compare density-01 device regimes. [data agent already doing this]
  2. Gap/zero-inflation analysis: per-sensor missingness, zero share, runs of zeros, correlation with battery/LoRaWAN.
  3. Cross-sensor correlation: at matched timestamps; pairwise correlation; effective number of independent sensors; coefficient of variation. This tests whether the “8-sensor field” is really 8 independent measurements.
  4. Subsampling experiment: train the existing global model on 1-in-k rows (simulate 30 min, 1 h, 2 h cadence) and measure MAE/WAPE degradation. This quantifies the value of temporal sampling and supports the “increase duty cycle or use continuous probes” recommendation.
  5. Leave-one-sensor-out: train on 7 sensors, test on the held-out sensor. If the global model fails, the “one model for all sensors” claim is weak; if it succeeds, it supports shared structure.
  6. Factor sensitivity: rescale the counts by 0.5/0.7/1.0/1.4 and compare the staffing plan (coverage/cost) under each. This shows why calibration matters for decisions, even if relative patterns are stable.
  7. Decision-level evaluation: compare the staffing plan to a flat schedule on coverage and cost under explicit assumptions; report it as a heuristic comparison, not ROI.

These experiments can be done in paper/research/ with the existing venv. They are optional for the hackathon report but strengthen the extended paper.

7. What to keep from the hackathon build

  • Decision-first framing: the plan, not the dashboard.
  • Global model across sensors; time-based holdout; seasonal-naive baseline.
  • Transparent staffing heuristic with visible assumptions.
  • Cached fallback and staleness labels.
  • Build-time batch for a 24-hour demo.
  • Honest limitations and non-claims.

8. What to replace or add for the “right way”

Layer Hackathon Right way Trigger
Sensing 8 BLE PAX counters Thermal at doorways; mmWave at wide sites; 2–4 cameras as reference; probe data for city context; manual counts for calibration Before any absolute-count or ROI claim
Calibration Fixed 0.7 factor Quarterly manual counts; capture-recapture/penetration study; Bayesian calibration with uncertainty Before staffing cost claims
Privacy Aggregate counts; no raw MACs in app Edge aggregation; local DP/secure aggregation; k-threshold; no raw trajectories; DPIA; legal basis; retention limits Before public deployment
Spatial Local pulse map Network-aware negative-binomial hierarchical field at 100–150 m; uncertainty/reliability mask Before heatmap-based decisions
Forecast Global HistGB; point + band Quantile GBM; conformal calibration; forecast-weather evaluation; decision-level metrics Before operational use
Platform 3 pods; tunnel; cached fallback Auth/SSO; observability; SLOs; multi-tenant; IaC; cost controls When external users/uptime matter
Evaluation MAE/WAPE/R²; plan latency Ground truth; calibration; decision regret; privacy audit; A/B or pilot Before claiming ROI or generalization

8.5 Tone and positioning

The paper should be constructive, not adversarial. Code for Heilbronn is a volunteer civic-tech group that built a low-cost, privacy-conscious prototype under real constraints. The critique is of the claim (“privacy-preserving,” “footfall,” “15-minute counts”) and of the measurement design, not of the volunteers’ intent. Acknowledge:

  • the genuine data-minimisation choices (no raw MAC storage, counts only over LoRaWAN, locally administered filter);
  • the hardware/power constraints that motivate a 10 s scan and a 15 min sleep;
  • the value of the network as a relative activity index for local decisions;
  • the fact that the challenge brief, not the firmware alone, frames the data as “people.”

The paper’s constructive contribution is the measurement model, the calibration program, and the privacy-first alternative. Avoid “the sensor is fake” or “the data are worthless.” The defensible claim is: the data are useful for relative patterns in a local cluster, but not for absolute footfall, city-wide heatmaps, or ROI claims without calibration.

9. Open decisions for the human

  1. Paper shape: extended 10–12 page Zenodo paper (recommended) vs 4–5 page hackathon report vs both.
  2. New experiments: run the subsampling/LOSO/factor-sensitivity analyses? (Recommended: yes, post-demo, label as analysis.)
  3. Prototype: design/architecture paper only, or a small new prototype? (Recommended: design + evaluation plan; no new build before the demo.)
  4. Legal scope: Germany/EU focus with a short generalization note? (Recommended: yes.)
  5. Publication: Zenodo DOI first, arXiv if endorsement? (Recommended: yes.)
  6. Reference budget: 25–35 refs for the extended paper; 8–12 for the hackathon report. Confirm the deviation from agents/paper.md is acceptable.

10. Source notes

  • 01-bluetooth-pax-reliability.md — BLE/PAX reliability, code trace, forensic quantization, literature.
  • 02-privacy-law-mac.md — pending; legal classification, ePrivacy/TTDSG, DPIA, enforcement.
  • 03-google-maps-probe-data.md — Google Maps traffic/ETA, terminology, privacy model, alternatives.
  • 04-sensing-alternatives.md — camera/thermal/radar/LiDAR/CSI/telco/GPS/transit/manual/fusion.
  • 05-heatmap-forecasting.md — sparse-to-heatmap, forecasting, covariates, transfer, evaluation.
  • 06-applications-matrix.md — six challenges, shared abstraction, transferability, case studies.
  • 07-data-quality.md — pending; empirical data-quality analysis and figures.
  • 08-privacy-enhancing-technologies.md — DP, k-anonymity, secure aggregation, synthetic data, threat models.
  • 09-paxcounter-code-audit.md — our source-code audit of the deployed lorax firmware and upstream reference.