Singling out
Singling out is a fundamental criterion in evaluating the anonymity of synthetic data. It refers to the risk that an individual can be uniquely identified within a dataset based on the available information.
In current privacy literature, this mechanism is often described as record isolation. In practice, it refers to the same idea as singling out: an attacker can isolate a row or a rare combination of attributes within a dataset. The criterion becomes relevant only when that isolated record can be linked to a specific physical individual using reasonably available background knowledge.
In other words, isolating a rare attribute combination in a table is not, by itself, a privacy failure if it cannot be associated with a real person. The risk becomes material when the isolated combination can realistically be tied to one person.
Consider a dataset containing employee information within an organization. An attacker who knows an employee’s department, age, and seniority could use this information to isolate the individual, especially if these attributes make them unique—for example, being the only employee with more than 10 years of seniority in a small team.
List of metrics measuring singling out criteria
- Hidden rate
- Local Cloaking
- Distance to closest
- Closest distances ratio
- Closest Rate
- Column direct match protection
- Row direct match protection
- Membership Inference Attack
Factors Influencing the Risk of Singling Out
The risk of singling out increases proportionally with the amount and precision of information available to an attacker. The more contextual data (e.g., age, location, profession) an attacker possesses, the more likely they are to isolate an individual in the dataset.
This criterion is particularly critical for individuals with remarkable values (outliers), meaning rare or atypical characteristics. For instance, someone living in a sparsely populated area or having a unique combination of traits in the dataset is far more vulnerable to singling out.
Importance of Mitigation
To limit the risks of singling out, the metrics we have developed systematically assess the likelihood that an attacker, with varying levels of prior knowledge, could isolate an individual. These analyses help identify and address vulnerabilities before the data is used or shared.