Anonymization

Anonymization transforms personal data to protect individual privacy. This project relies on the RECITALS framework with the ARX backend to anonymize the Adult dataset.

---

Role of this step

Anonymization is the starting point of the pipeline. It produces a transformed dataset on which all utility evaluations (ARX metrics and classification) are then computed.

---

Dataset: Adult (UCI)

The Adult dataset contains 30,000+ records describing individuals from the U.S. Census.

Quasi-identifiers (QI)

Quasi-identifiers are attributes that, when combined, may allow re-identification of an individual.

Attribute	Type	Example
age	Numeric	39
sex	Categorical	Male, Female
race	Categorical	White, Black, Asian-Pac-Islander…
marital-status	Categorical	Married-civ-spouse, Never-married…
native-country	Categorical	United-States, Mexico…

Sensitive attribute

Attribute	Type	Values
income	Binary	<=50K, >50K

---

Generalization hierarchies

For each quasi-identifier, a hierarchy defines the possible levels of generalization — from most specific to most general.

flowchart TD
    A["39"] --> B["35–39"]
    B --> C["30–39"]
    C --> D["*"]

• Numeric attributes (age): grouped into intervals (35–39, 30–39, …)
• Categorical attributes (sex, race, …): grouped into broader categories up to * (full suppression)

Hierarchies are defined as CSV files in the hierarchies/ folder (one file per attribute).

---

Privacy models

k-anonymity

Each record is made indistinguishable from at least k − 1 other records on the quasi-identifiers. The re-identification risk is bounded by 1/k.

\[\forall \text{ equivalence group } G,\quad |G| \geq k\]

l-diversity

Extension of k-anonymity: each equivalence group must contain at least l distinct values for the sensitive attribute, limiting inference on that attribute.

\[\forall \text{ equivalence group } G,\quad |\{s \in G\}| \geq l\]

Suppression limit

A suppression_limit parameter (in %) allows ARX to remove a fraction of records when it is impossible to anonymize them while satisfying privacy constraints. Default: 10%.

---

Benchmark pipeline

Experiments are automatically generated from a parameter grid.

flowchart TD
    A[benchmark_grid.json\nParameter grid] --> B[run_benchmark.py]
    B --> C[Config generation\none per combination]
    C --> D[run_one_experiment.py]
    D --> E[outputs/anonymized/\nAnonymized data]
    D --> F[outputs/metrics/\nARX metrics]
    D --> G[benchmark_summary.csv\nSummary]

Grid parameters

Parameter	Explored values
QI subset sizes	2, 3, 4, 5
k (k-anonymity)	2, 5, 10, 20
l (l-diversity)	2
Suppression limit	10%

Each combination generates an independent experiment with its own identifier, JSON configuration, and output files.