Overview

This project is part of a TER (Travail d'Étude et de Recherche) and studies the impact of data anonymization on data utility — both from a statistical standpoint and from a supervised classification standpoint.

---

Objective

Anonymization protects individual privacy by transforming data before publication. This transformation comes at a cost: it degrades the available information. The objective of this project is to quantify the privacy-utility trade-off, using the Adult (UCI) dataset and the RECITALS anonymization framework (ARX backend).

---

Dataset

The dataset used is the Adult dataset (UCI Machine Learning Repository), containing 30,000+ records describing individuals from the U.S. Census.

The target attribute is income (income): <=50K or >50K.

---

General pipeline

flowchart TD
    A[adult.csv\nOriginal data] --> B[Anonymization\nRECITALS / ARX]
    B --> C[Anonymized data]
    C --> D[ARX Utility Metrics\nQuality models · Aggregate functions]
    C --> E[Classification\nZeroR · LR · NB · RF]
    D --> F[Utility evaluation]
    E --> F

---

Project structure

Section	Description
Anonymization	Privacy models, generalization hierarchies, benchmark pipeline
Quality Models	ARX quality models used to guide optimization
Aggregate Functions	Aggregating per-attribute scores into a global score
ARX Utility Metrics	Post-anonymization analysis metrics from RECITALS
Classification	Utility evaluation through supervised classifiers

---

Tools

• RECITALS — Python anonymization framework, ARX backend
• ARX — Java anonymization engine (k-anonymity, l-diversity)
• scikit-learn — classification pipelines
• pandas / numpy — data processing