Hugdrigo research labs / project 75

Privacy intelligence

Measure disclosure risks, explore anonymization and audit the trade-offs.

Disclosure surface

Baseline and transformed equivalence groups, recalculated from source records.

Equivalence class distribution

Number of records belonging to groups of each size. Smaller groups are easier to single out.

Transformed retained records

Privacy profile

Check exposure, retention and generalization side by side.

Recommended next step

Generalization & suppression

Configure transformations for quasi-identifiers and inspect actual retained records.

Transformation controls

Direct identifiers are never included in anonymized CSV downloads. Suppression removes entire records from export when their equivalence group fails either target.

Preview & quality checks

Preview (first 8 retained rows)

Equivalence class explorer

Every class is defined by the current generalized quasi-identifiers.

Distinct l-diversity counts different values of the sensitive outcome per group. A group that meets k and l may still leak information about an individual.

Privacy optimization laboratory

Exhaustively evaluates all 36 global generalization combinations under the selected privacy targets.

Optimization limits

Suppression allowance is a percentage of the original dataset.
Higher values strongly discourage removal of records.

Search coverage

Every combination is assessed deterministically. Ranking minimizes weighted generalization loss plus a penalty for suppressed records. This is an exact search of the stated 36 options, not an optimality claim over all possible anonymization methods.

Top feasible configurations

Utility & representation

Measure lost records, information coarsening and possible distortion of aggregate statistics.

Outcome composition

Outcome distribution before and after suppression. Different proportions can signal selection bias.

Information preserved

Illustrative attribute-level fidelity score, not a guarantee of research utility.

The age variable is only generalized. IDs are removed entirely. Averages computed on retained records can differ from original averages due to suppression.

Methodological limits

k-anonymity limits unique quasi-identifier combinations within the released table. Distinct l-diversity requires at least l distinct outcomes in each retained equivalence class. Neither offers a formal guarantee of differential privacy. The tool does not model attacker background knowledge, linkage to external records, t-closeness, membership disclosure or inferential attacks. Use only de-identified, authorized records for demonstration and obtain specialist review for any real disclosure.

Data workbench & audit exports

Import local data and export transformed results or a reproducible configuration.

Browser-only processing

Import CSV

Required columns: id, age, region, role, outcome. Maximum 5,000 rows and 1 MB. Existing data remains intact if validation fails.

Imported data is read locally; no upload request is made by this application. However, avoid placing actual personal or confidential information into a public-facing demo application. Changes are not saved between sessions unless exported.

Downloads

Generate local, reproducible research outputs. Anonymized CSV intentionally excludes direct IDs and suppressed rows.

Source data checks

Current dataset fields

PrivacyForge v1.0 · Hugdrigo · Project #75 · MIT License · Research and educational demonstration only · No external libraries, analytics or API calls.