When Poor-Quality Data Meet Anonymization Models: Threats and Countermeasures

Abstract

In the modern era, the avenues for data generation have significantly evolved, and therefore, large-scale and diverse types of data are being collected for downstream tasks. However, some automated tools can curate data with many vulnerabilities (e.g., missing values, wrong values, outliers, skewed distributions, missing labels, etc.) which can hamper its usage in underlying applications. For example, skewed data may lead to imbalanced learning when used in machine learning (ML) classifiers. Similarly, low-quality data can inadvertently propagate bias in ML decisions, leading to conflict or disputes. Data quality enhancement with the least possible cost has become a hot research area. In this paper, we demonstrate how poor-quality data can pose serious threats to anonymization models, consequently undermining privacy and utility requirements. We propose various countermeasures to inspect and improve data quality before anonymization so as to not lose guarantees of both privacy and utility. Specifically, we pinpoint eleven different threats from a small segment of data to underscore the relevance and urgency of such issues in the anonymization domain. We devise six practical countermeasures to provide resilience against these potential threats to enhance the performance and resilience of anonymization models. We aim to uncover the ways poor-quality datasets are handled by anonymization models, and what threats to both privacy and utility exist. Our work can guide the privacy and database community to improve the mainstream technologies used for privacy preservation to effectively resolve present-day privacy threats. To the best of the authors’ knowledge, this is the first work that highlights the threats posed by the poor quality data to the popular anonymization models and suggests countermeasures to overcome them along with reasonable experiments on two datasets.