Autoencoder-Enhanced Hierarchical Mondrian Anonymization via Latent Representations.

Area of Science:

• Data privacy and security
• Information science
• Computer science

Background:

• Releasing structured microdata presents challenges in balancing data utility with privacy protection against group-based disclosure risks.
• Existing anonymization techniques often struggle to effectively manage sensitive attribute concentration and diversity within anonymized data groups.

Purpose of the Study:

• To introduce AE-LRHMA, a hybrid anonymization framework designed to enhance data utility while mitigating privacy risks.
• To provide explicit control over sensitive value concentration and diversity within equivalence classes.
• To evaluate the performance of AE-LRHMA against established anonymization baselines.

Main Methods:

• AE-LRHMA employs Mondrian-style hierarchical partitioning within an autoencoder-learned latent space.
• The framework integrates local (k,e)-microaggregation techniques.
• A tunable constraint set, including k, a maximum sensitive proportion threshold, and an optional sensitive-entropy threshold, is introduced for fine-grained control.

Main Results:

• Experiments on the Adult and Bank Marketing datasets show AE-LRHMA achieves lower information loss compared to baselines.
• The framework demonstrates more stable group structures in the anonymized data.
• Linkage-attack-oriented risk metrics indicate favorable relative disclosure trends.

Conclusions:

• AE-LRHMA offers an effective hybrid approach for structured microdata anonymization.
• The method provides a tunable mechanism to balance data utility and privacy, outperforming existing techniques.
• Further empirical analysis supports its utility in managing disclosure risks.