Skip to content

File Formats

This page describes every file involved in the pipeline — inputs, configuration files, generalization hierarchies, and all generated outputs.

Source data

data/adult.csv

The original UCI Adult dataset. This file is never modified — it serves as the reference throughout the entire pipeline.

Field Description
Format CSV, comma-separated
Rows ~30,000 records
Columns 14 attributes (age, workclass, education, marital-status, occupation, relationship, race, sex, capital-gain, capital-loss, hours-per-week, native-country, income)
Role Ground truth for all utility comparisons

data/adult_with_record_id.csv

A copy of adult.csv with an added record_id column (unique integer per row). Used by linkage and MIA attack scripts that need to track individual records across transformations.

Configuration files

configs/base_config.json

Base configuration shared across all experiments. Defines the dataset path, quasi-identifiers, sensitive attribute, and default privacy parameters.

{
  "dataset": "data/adult.csv",
  "quasi_identifiers": ["age", "sex", "race", "marital-status", "native-country"],
  "sensitive_attributes": ["income"],
  "k": 5,
  "l": 2,
  "suppression_limit": 10,
  "utility_measure": "loss",
  "utility_aggregate": "arithmetic_mean",
  "backend": "arx"
}

configs/benchmark_grid.json

Defines the parameter grid used by run_benchmark.py to generate all experiment combinations.

{
  "qi_subset_sizes": [2, 3, 4, 5],
  "k_values": [2, 5, 10, 20],
  "l_values": [2],
  "suppression_limits": [10]
}

Each combination of parameters generates one independent experiment.

configs/classification_config.json

Configuration for the classification utility benchmark. Specifies the number of folds, the target attribute, the active models, and their hyperparameters.

{
  "n_folds": 10,
  "random_state": 42,
  "target": "income",
  "active_models": ["naive_bayes"],
  "feature_cols": ["age", "sex", "race", "marital-status", "native-country"]
}

All parameters — including hyperparameters for each model — are fully configurable in this file.

outputs/configs/{experiment_id}.json

Runtime configuration generated for each experiment. It is a resolved copy of base_config.json with all paths made absolute, the hierarchy mapping filled in, and the exact parameters used for that run. Serves as a reproducibility record.

Generalization hierarchies

hierarchies/{attribute}.csv

One CSV file per quasi-identifier. Each row defines one value and its successive generalizations, from the most specific (level 0) to the most general (usually *).

Example — hierarchies/age.csv:

25;25-29;20-29;*
26;25-29;20-29;*
27;25-29;20-29;*
...
Column index Meaning
0 Original value
1 Level 1 generalization
2 Level 2 generalization
… …
Last Full suppression (*)

Available hierarchy files: age.csv, sex.csv, race.csv, marital-status.csv, native-country.csv, education.csv, occupation.csv, workclass.csv, relationship.csv.

Anonymization outputs

outputs/anonymized/{experiment_id}.csv

The public anonymized dataset produced by ARX for a given experiment. Quasi-identifier values are replaced by generalized values or * (suppressed). The sensitive attribute (income) is kept unchanged.

This is the file used as input for all utility evaluations.

outputs/anonymized_eval/{experiment_id}.csv

Internal version of the anonymized dataset that retains additional columns used during evaluation (e.g. fold indices, record identifiers). Not published — used only by evaluation scripts.

outputs/metrics/{experiment_id}.json

ARX result metrics collected after each anonymization. Contains:

  • Equivalence class statistics (count, min/avg/max size, suppressed records)
  • Per-attribute metrics (granularity, non-uniform entropy, generalization intensity, squared error)
  • Global metrics (discernibility, ambiguity, SSESST, record-level squared error)
  • Optimization scores (min and max)
  • Anonymization time and number of transformations explored

outputs/benchmark_summary.csv

Master summary table — one row per experiment. Aggregates key parameters and metrics from all completed experiments. Used as the entry point by classification and reporting scripts.

Column Description
experiment_id Unique experiment identifier
status success or error
quasi_identifiers Pipe-separated list of QI used
k, l, t Privacy model parameters
suppression_limit Max suppression percentage
csv_path Path to the anonymized CSV
config_path Path to the runtime config
+ ARX metrics All metrics from outputs/metrics/

Classification outputs

outputs/classification/{experiment_id}_classification.json

Full classification results for one experiment. Contains, for each model (ZeroR, LR, NB, RF):

  • Global accuracy (input and output)
  • Relative accuracy
  • Per-class sensitivity and specificity
  • AUC and ROC curve points (100 sampled)
  • Brier score and Brier skill score

outputs/classification_summary.csv

Flattened summary of classification results across all experiments — one row per experiment per model. Used for cross-experiment comparison.

File classification

File type Perspective Usage
data/adult.csv Reference Never modified — ground truth
configs/*.json Configuration Define experiments and models
hierarchies/*.csv Input Define generalization trees
outputs/anonymized/*.csv Public output Result of anonymization
outputs/metrics/*.json Internal ARX analysis metrics
outputs/benchmark_summary.csv Internal Experiment index
outputs/classification/*.json Internal Classification evaluation results
outputs/configs/*.json Internal Reproducibility record