New threats to health data privacy

Fengjun Li, Xukai Zou, Peng Liu, Jake Y. ChenBMC Bioinformatics 2011 · DOI

TL;DR

This paper highlights a critical gap in health data privacy: current de-identification methods are insufficient against sophisticated attacks that leverage publicly available online information. The authors demonstrate how disparate data from social networks and public records can be aggregated to re-identify individuals and expose sensitive health information, even with approximate or inaccurate initial data.

What Problem Does This Paper Address?

The increasing digitalization of health records (EHRs) promises improved healthcare efficiency and research capabilities. However, this also raises significant privacy concerns, especially when data needs to be published for secondary use. Existing research primarily focuses on de-identifying or anonymizing data to remove direct identifiers. The authors argue that this approach is insufficient because it overlooks the vast amount of personal information voluntarily shared online through social networks and readily available public records. These external sources, combined with advanced search and data aggregation techniques, empower adversaries to link seemingly innocuous pieces of information and reconstruct detailed profiles, ultimately compromising patient privacy. The paper specifically addresses the vulnerability of health data to attacks that combine information from multiple online sources, a threat that has received limited attention.

Methods

The paper's methodology centers on demonstrating the feasibility and effectiveness of privacy attacks through a combination of attribution, inference, and aggregation.

  • Attribution: This involves extracting identifiable or semi-identifiable attributes from various online sources. The paper highlights the use of online social networks (OSNs) and public databases (e.g., voter registration, phone books) as primary sources.
  • Inference: This stage uses statistical learning or logical reasoning to discover additional attributes from the extracted information and social activities.
  • Aggregation: This is the core of the attack, where records from different sources that potentially belong to the same individual are linked. The paper emphasizes that this linking can be achieved through strong evidence (matching identifiers or quasi-identifiers) or weak evidence (statistical similarities).

A key novel aspect is the exploration of attacks using approximate information. The authors introduce the concept of "k-approximate-anonymity" to formalize this, defining it as a state where a record cannot be distinguished from k-1 other records within a certain distance threshold (l). This contrasts with traditional k-anonymity, which assumes exact matches.

Datasets

The paper uses a simulated real-world case study involving a hypothetical patient named "Jean."

  • Simulated Data: The primary "dataset" is constructed through a simulated attack scenario. This involves crawling profiles from two online medical social networking sites (MedHelp and MP and Th1 Discussion Forum) and then using web search engines and public data sources (phone books, birth records, etc.) to gather additional information.
  • Publicly Available Data: The authors mention using a residential phone book from State College, PA, containing 24,399 records, for their k-approximate-anonymity experiments. This dataset included phone numbers, names, and residential addresses.

The paper does not introduce a new benchmark dataset but rather uses existing public data and simulated online profiles to illustrate attack vectors.

Key Insights and Ideas

  • The "Information Rich" Context: The digital age, with its abundance of voluntarily shared online data, creates a perfect storm for privacy breaches, making individuals highly vulnerable even with seemingly innocuous information.
  • Beyond Direct Identifiers: Traditional de-identification methods that focus solely on removing direct identifiers (like names or SSNs) are insufficient. The real threat lies in the aggregation of quasi-identifiers and semi-identifiable attributes from disparate sources.
  • The Power of Approximate Information: Adversaries don't need perfect information. Even with inaccurate or incomplete data (e.g., a misspelled name), sophisticated linking techniques can still lead to successful re-identification. This is a significant departure from many privacy models that assume perfect knowledge.
  • Social Networks as Goldmines: Online social networks, where users willingly share personal details, are prime targets for attackers. They provide a rich source of semi-identifiable attributes that can be linked to other data.
  • The "Attribution, Inference, Aggregation" Trifecta: This three-step process is a powerful framework for understanding how privacy can be compromised in the modern online landscape.

Key Results

  • Successful Re-identification in Case Study: The simulated attack on "Jean" successfully recovered a comprehensive profile including her full name, date of birth, husband's name, home address, phone numbers (home and cell), two email addresses, occupation, and detailed medical information (lab test results). This was achieved by linking profiles from two medical social networks with public records and search engine results.
  • Vulnerability to Approximate Information: The k-approximate-anonymity experiments demonstrated that even with an edit distance of l=2 (allowing for two character differences), over 70% of individuals in a sample population were still uniquely identifiable. This highlights the significant risk posed by attackers with imperfect information.
  • High Identifiability: The study suggests that people are highly identifiable even with inaccurate information, especially when combined with readily available external data sources.

Limitations and Future Work

  • Simulated Attack: The primary case study is a simulation. While it demonstrates the potential for such attacks, it doesn't represent a real-world breach with all its complexities.
  • Focus on Specific Data Sources: The study primarily focuses on social networks and public records. Other emerging data sources (e.g., IoT devices, wearable tech) are not explicitly covered.
  • Theoretical Framework for Approximate Privacy: While k-approximate-anonymity is introduced, its practical implementation and effectiveness across diverse datasets and attack scenarios warrant further investigation.
  • Future Work: The authors suggest the need for more robust privacy protection tools that can defend against these sophisticated aggregation and inference attacks. They also imply that a deeper understanding of the types and amounts of external knowledge adversaries can access is crucial for developing effective data sanitization strategies.

Summary updated

BibTeX
@article{li2011new,
  author = {Fengjun Li and Xukai Zou and Peng Liu and Jake Y. Chen},
  title = {{New threats to health data privacy}},
  journal = {BMC Bioinformatics},
  year = {2011},
  doi = {10.1186/1471-2105-12-s12-s7},
}
Figure