Systematically investigate the relationship between record uniqueness and membership-inference vulnerability

Investigate the relationship between the re-identification risk associated with unique attribute combinations and vulnerability to membership-inference attacks, including whether records that are more susceptible to re-identification are also more vulnerable to membership inference.

Background

The paper discusses k-anonymity as a preprocessing defense that makes each record indistinguishable from at least k−1 other records. Unique attribute combinations increase re-identification risk and may also increase membership leakage because models can exert greater influence on, or memorize, rare records.

Although the paper later empirically studies single-out vulnerability, it explicitly identifies the broader relationship between uniqueness, re-identification risk, and membership-inference vulnerability as not having been systematically investigated. Establishing this relationship would help determine whether k-anonymity and related anonymization techniques provide reliable protection against membership inference, particularly for records with rare quasi-identifier combinations.