Data Anonymization.pdf

Data Anonymization as a Vector Quantization Problem:
Control Over Privacy for Health Data is a research paper that tackles data anonymization from a vector quantization point of view. The goal is to anonymize data to avoid single individual or group re-identification while maintaining data integrity and structure.

The paper proposes using vector quantization to capture the structure of the data through clustering, and then anonymizing the data by "pushing back outliers in the crowd" and modifying the data to blend in individuals while preserving cluster-level statistics and structure.

The approach involves considering data fields as separate entities and building clusters of records based on metric proximity. The assumption is that if a group is large enough and the records inside have been "stirred" enough, identification of a single individual becomes impossible.

The paper uses an example of health-related information to illustrate the concept, where relationships between non-sensitive fields can make it easy to identify individuals. The proposed approach seeks to "blend in" individuals with the rest of the records and make the data anonymized by modifying the data values to realistic ones while preserving some of the information.

The high-level motivation for data privacy is to distort the data in a known manner to increase privacy or entropy, without changing the format of the data or the underlying space. The paper aims to propose practical solutions to the identification problem by defining notations and assumptions on the structure of the data and tackling the problem of moving single samples or groups back into bigger clusters.

Key points:
- Data anonymization through vector quantization
- Clustering to capture data structure
- Anonymizing data by "pushing back outliers" and modifying data to blend in individuals
- Preserving cluster-level statistics and structure
- Example of health-related information to illustrate the concept
- High-level motivation for data privacy to distort data without changing format or underlying space.