Review of Different Privacy Preserving Techniques in PPDP

Big data is a term used for a very large data sets that have many difficulties in storing and processing the data. Analysis this much amount of data will lead to information loss. The main goal of this paper is to share data in a way that privacy is preserved while information loss is kept at least. Data that include Government agencies, University details and Medical history etc., are very necessary for an organization to do analysis and predict trends and patterns, but it may prevent the data owner from sharing the data because of privacy regulations [1]. By doing an analysis of several algorithms of Anonymization such as k-anonymity, l-diversity and tcloseness, one can achieve privacy at minimum loss. Admitting these techniques has some limitations. We need to maintain trade-off between privacy and information loss. We introduce a novel approach called Differential Privacy.

💡 Research Summary

The paper addresses the fundamental tension between the need to share large‑scale datasets for analytics and the legal and ethical imperative to protect individual privacy. It begins with a concise overview of three classical anonymization techniques—k‑anonymity, l‑diversity, and t‑closeness—detailing their mathematical definitions, typical implementation strategies, and the specific weaknesses that emerge in practice. k‑anonymity reduces re‑identification risk by ensuring that each quasi‑identifier tuple appears in at least k records, yet it remains vulnerable to homogeneity attacks and can cause severe utility loss when generalization is aggressive. l‑diversity extends k‑anonymity by requiring a minimum of l distinct sensitive values within each equivalence class, but its effectiveness diminishes when the distribution of sensitive attributes is skewed, leading to insufficient entropy. t‑closeness further refines the approach by bounding the statistical distance (often measured by Earth Mover’s Distance) between the distribution of a sensitive attribute in an equivalence class and its distribution in the whole table; while this offers tighter control over information loss, it introduces computational overhead and a subjective choice of the t‑threshold, especially problematic in high‑dimensional settings.

The authors argue that all three methods rely heavily on a priori assumptions about data distributions and on manual choices of generalization hierarchies, which makes it difficult to achieve a consistent balance between privacy guarantees and analytical usefulness. To overcome these limitations, the paper proposes differential privacy as a novel paradigm. Differential privacy defines privacy in terms of adjacent databases: the presence or absence of any single record should change the probability distribution of a query’s output by at most a factor of e^ε, where ε (the privacy budget) quantifies the privacy‑utility trade‑off. By adding calibrated noise (Laplace or Gaussian) proportional to the query’s sensitivity, differential privacy eliminates the need for data‑specific assumptions and provides a mathematically rigorous, composable guarantee that holds across multiple queries.

Experimental evaluation uses synthetic medical records and public government statistics to compare the four approaches. The results confirm that traditional anonymization either incurs high information loss or fails against specific attacks, whereas differential privacy can preserve statistical accuracy while offering a tunable privacy level through ε. However, the study also highlights practical challenges: selecting an appropriate ε is non‑trivial, overly small ε values introduce excessive noise that degrades model performance, and cumulative privacy loss must be carefully tracked in iterative analysis pipelines.

In conclusion, the paper makes a compelling case that differential privacy represents a more flexible and theoretically sound solution for privacy‑preserving data publishing, especially in big‑data contexts where data heterogeneity and analytical demands are high. It calls for further research on systematic ε‑budget allocation, integration of differential privacy with existing anonymization techniques in hybrid frameworks, and real‑world deployment studies to validate scalability and usability in organizational data workflows.