geomask

Geographic masking of confidential point data, with the privacy and utility metrics you need to document the trade-off.

Public-health, social-science and urban researchers often hold point data that identify people: home addresses of patients, locations of incidents, survey respondents' dwellings. Publishing those points, or even a detailed map of them, can let anyone with a street map work out who is who. geomask implements the established geomasking methods from the literature and, just as importantly, the metrics that tell you how much privacy a mask buys and how much analytical value it costs.

• Masks: random perturbation, Gaussian displacement, donut masking, population-adaptive donut masking with a verified spatial k-anonymity guarantee, grid and hexagon snapping, grid / hexagon / Voronoi / label aggregation with small-cell suppression, location swapping and adaptive areal elimination (AAE).
• Privacy metrics: spatial k-anonymity per point, expected re-identification risk, nearest-neighbour linkage rate, displacement statistics.
• Utility metrics: Ripley's K / L difference, kernel-density correlation, Moran's I and hot-spot overlap of gridded counts, mean displacement.
• report(): a privacy-vs-utility table across methods.
• Reproducible (every random mask takes a seed), pure NumPy (SciPy optional for KD-tree speed), works in metres with built-in lon/lat helpers (UTM and local equirectangular), typed, and a CLI.

Masking is not anonymisation by itself. Read the ethics and limitations section before you share anything.

Installation

pip install geomask            # NumPy only
pip install "geomask[fast]"    # + SciPy KD-trees for large data sets

Python 3.9 or later.

Quick start

All examples use synthetic data generated by geomask.synthetic_city; no real person is represented anywhere in this project.

import geomask as gm

town = gm.synthetic_city(n_population=5000, n_cases=300, seed=1)
cases, homes = town.cases, town.population   # (n, 2) arrays in metres

# Donut mask: every point moves between 50 m and 300 m.
masked = gm.donut_mask(cases, min_radius=50, max_radius=300, seed=1)

# Population-adaptive donut: radii grow in sparse areas and every masked point
# is re-drawn until at least 10 households are closer to it than the true home.
safe = gm.donut_mask(cases, 50, 300, seed=1, population=homes, k=10)

k = gm.spatial_k_anonymity(cases, safe, homes)
print(k.min())                          # >= 10, audited independently

gm.report(gm.evaluate(cases, safe, population=homes, k=10))

Comparing methods

examples/compare_methods.py masks the same synthetic cases with every method and prints the trade-off table:

results = gm.compare(cases, {
    "perturb r=250": gm.random_perturbation(cases, 250, seed=1),
    "gaussian s=100": gm.gaussian_displacement(cases, 100, seed=1),
    "donut 50-300": gm.donut_mask(cases, 50, 300, seed=1),
    "donut k>=10": gm.donut_mask(cases, 50, 300, seed=1, population=homes, k=10),
    "swap r=300": gm.location_swap(cases, homes, 300, seed=1),
    "AAE k=20": gm.adaptive_areal_elimination(cases, homes, k=20, cell_size=100, seed=1),
    "grid 250 k=5": gm.aggregate_to_grid(cases, 250, k=5).masked_points,
    "hex 150 k=5": gm.aggregate_to_grid(cases, 150, k=5, shape="hex").masked_points,
}, population=homes, k=10)
gm.report(results)

method          k min  k med   <k %   risk  NN re-id  disp m  min m  dL max m  KDE r  Moran I  hot J  suppr
--------------  -----  -----  -----  -----  --------  ------  -----  --------  -----  -------  -----  -----
perturb r=250       0     42  20.0%  0.081     19.7%     165     19      26.9  0.999    0.700   0.36      0
gaussian s=100      0     19  35.0%  0.174     34.7%     118      4      23.1  0.998    0.702   0.47      0
donut 50-300        0     62  13.7%  0.049     15.7%     203     55      37.2  0.998    0.738   0.67      0
donut k>=10        10     66   0.0%  0.023      8.7%     224     66      31.9  0.997    0.755   0.62      0
swap r=300          1     88  12.0%  0.035     13.3%     207     26      45.3  0.989    0.663   0.53      0
AAE k=20            0     24  27.0%  0.122     32.3%     238      6      45.2  0.993    0.678   0.55      0
grid 250 k=5        0     18  28.5%  0.106      6.3%      93     14     727.8  0.932    0.507   0.25    170
hex 150 k=5         0     18  30.1%  0.120      6.7%      93     21     668.0  0.945    0.538   0.33    167
n = 300; k threshold = 10; Moran I of original counts = 0.708

How to read it: only the population-adaptive donut guarantees that every point has at least k = 10 households closer to its released location than the true one (k min), while keeping the density surface almost unchanged (KDE r 0.997). Fixed-radius masks leave some rural points weakly protected (<k %). Aggregation with suppression drops more than half the points in this small sample (suppr) and distorts second-order structure (dL max). AAE's guarantee is areal (each release area holds at least k residents), which the point-based k min column does not measure - see below.

Methods

Notes on parameters:

• Coordinates are projected, in metres. Use gm.make_projection(lonlat) (UTM zone of the data by default, or method="local") and its forward / inverse methods to convert.
• population is any set of locations that people could plausibly live at: address points, building centroids or the centres of a population raster. Repeat a centre once per resident to weight it (AAE also accepts population_weights). The true location is usually in this set; that is expected and handled.
• donut_mask(..., distribution="distance") reproduces the original donut formulation (uniform distance); the default "area" is uniform over the annulus, which puts fewer points close to the inner ring.
• Aggregation results keep the true counts in counts; publish released_counts or table(), which show suppressed cells as <k.

Metrics

gm.evaluate() computes all of them; gm.compare() evaluates several masks with identical settings; gm.report() prints the table. Suppressed points (NaN rows) are excluded from the masked pattern and counted in suppr.

Command line

# Donut mask a lon/lat CSV (coordinates are auto-detected and projected to UTM)
geomask donut cases.csv --x lon --y lat --min 50 --max 300 --seed 1 -o masked.csv

# Population-adaptive donut with k >= 10
geomask donut cases.csv --x lon --y lat --min 50 --max 300 --k 10 \
    --population addresses.csv --seed 1 -o masked.csv

# Other methods
geomask perturb  cases.csv --radius 250 -o out.csv
geomask gaussian cases.csv --sigma 100 -o out.csv
geomask grid     cases.csv --cell 250 --k 5 -o out.csv
geomask hex      cases.csv --size 150 --k 5 -o out.csv
geomask voronoi  cases.csv --seeds clinics.csv --k 5 -o out.csv
geomask swap     cases.csv --radius 300 --population addresses.csv -o out.csv
geomask aae      cases.csv --k 20 --cell 100 --population addresses.csv -o out.csv

# Privacy / utility report (rows must be in the same order)
geomask evaluate cases.csv masked.csv --x lon --y lat --population addresses.csv --k 10
geomask evaluate cases.csv masked.csv --json

All other columns are copied unchanged; use --drop name address ... to remove direct identifiers. Suppressed points are written with empty coordinates. --coords lonlat|projected overrides auto-detection and --projection local uses an equirectangular projection instead of UTM.

Reproducibility

Every random function accepts seed (an int, None or a numpy.random.Generator). The same seed, input and version give identical output. Do not publish the seed together with the masked data: with the seed and the method parameters, an adversary who knows some true locations can test candidate locations. Store the seed with your confidential project records.

Ethics and limitations

• Masking is not anonymisation by itself. A masked point set can still be re-identified by linkage with other data (dates, ages, diagnoses, rare attributes), by repeated releases of differently masked versions of the same data, or by an adversary who knows the method and parameters. Treat masked data as reduced-risk personal data, not as anonymous data.
• Consult your IRB / ethics committee and data-protection officer (DPO) before sharing. They decide which risk is acceptable for your data, your participants and your recipients. This package helps you document the risk; it does not approve anything.
• GDPR note (EU/EEA and similar laws). Location data is listed as an identifier in Art. 4(1) GDPR, and health data are a special category (Art. 9). Under Recital 26, data remain personal data as long as a person is identifiable by means reasonably likely to be used; geomasked data for which you keep the original coordinates or a re-identification key are generally pseudonymised, not anonymised, and the GDPR still applies. A data-protection impact assessment (Art. 35) may be required. This is not legal advice.
• Guarantees are conditional. Spatial k-anonymity is only as good as the population layer: missing or outdated address data overstate k. AAE's guarantee is areal (at least k residents per release area), not a guarantee about distances.
• Small-cell suppression does not stop differencing attacks when several overlapping tables (other cell sizes, other years, sub-groups) are published, and it removes cases disproportionately from sparse, often rural or marginalised, areas - a bias to report in your methods.
• Edge effects: points near a study-area boundary, a coastline or an uninhabited area can be displaced to implausible places (except with swapping or AAE placement="population"), which an adversary can exploit to narrow the search.
• Utility metrics are global summaries. A high KDE correlation does not guarantee that a specific local cluster survived; re-run your actual analysis on the masked data before publishing results based on it.
• Use only synthetic or properly authorised data when experimenting. Never commit real personal data to a code repository.

API overview

Masks         random_perturbation, gaussian_displacement, donut_mask (DonutResult),
              location_swap, snap_to_grid, snap_to_hex,
              aggregate_to_grid, aggregate_to_voronoi, aggregate_by_label (AggregationResult),
              adaptive_areal_elimination (AAEResult)
Privacy       spatial_k_anonymity, expected_reidentification_risk,
              nearest_neighbour_reidentification_rate, displacement, displacement_stats
Utility       ripley_k, ripley_l, ripley_l_difference, kernel_density, kde_correlation,
              morans_i, cluster_preservation, mean_displacement
Evaluation    evaluate (EvaluationResult), compare, report
Projection    make_projection, UTM, LocalEquirectangular, utm_zone
Data          synthetic_city (SyntheticCity)
Backend       set_backend("numpy" | "scipy" | "auto"), get_backend

References

• Allshouse, W. B., Fitch, M. K., Hampton, K. H., Gesink, D. C., Doherty, I. A., Leone, P. A., Serre, M. L., & Miller, W. C. (2010). Geomasking sensitive health data and privacy protection: an evaluation using an E911 database. Geocarto International, 25(6), 443-452. https://doi.org/10.1080/10106049.2010.496496
• Armstrong, M. P., Rushton, G., & Zimmerman, D. L. (1999). Geographically masking health data to preserve confidentiality. Statistics in Medicine, 18(5), 497-525.
• Hampton, K. H., Fitch, M. K., Allshouse, W. B., Doherty, I. A., Gesink, D. C., Leone, P. A., Serre, M. L., & Miller, W. C. (2010). Mapping health data: improved privacy protection with donut method geomasking. American Journal of Epidemiology, 172(9), 1062-1069. https://doi.org/10.1093/aje/kwq248
• Kounadi, O., & Leitner, M. (2016). Adaptive areal elimination (AAE): a transparent way of disclosing protected spatial datasets. Computers, Environment and Urban Systems, 57, 59-67. https://doi.org/10.1016/j.compenvurbsys.2016.01.004
• Zandbergen, P. A. (2014). Ensuring confidentiality of geocoded health data: assessing geographic masking strategies for individual-level data. Advances in Medicine, 2014, 567049. https://doi.org/10.1155/2014/567049
• Zhang, S., Freundschuh, S. M., Lenzer, K., & Zandbergen, P. A. (2017). The location swapping method for geomasking. Cartography and Geographic Information Science, 44(1), 22-34. https://doi.org/10.1080/15230406.2015.1095655

Citing

If you use geomask in research, please cite it using the metadata in CITATION.cff and cite the original method papers above.

License

MIT (c) Yusuf Eminoğlu. See LICENSE.