Abstract
In sensitive environments like healthcare, data is increasingly stored in unstructured text formats such as medical reports and clinical notes. While auditing mechanisms for structured databases are well-established, monitoring keyword-based access to unstructured repositories remains a significant challenge. This paper introduces a novel auditing system designed to detect privacy violations by identifying suspicious queries that deviate from a user's professional profile. By integrating document relevance with topical interest modeling, the system provides a robust framework for post-hoc auditing in environments where rigid access controls are impractical.
Introduction
Organizations often face a "weakest link" problem in data security: while their structured databases may be protected by ironclad access controls and detailed SQL logs, their unstructured text repositories often remain under-monitored. In sectors like healthcare, traditional access control is frequently relaxed to ensure that medical professionals have immediate access to information during emergencies. However, this flexibility creates opportunities for "snooping"—unauthorized access to sensitive information for non-professional reasons.
Auditing unstructured data is inherently difficult because keyword queries are imprecise, and search results are returned as ranked lists rather than binary matches. This paper addresses these challenges by proposing a system that evaluates both the likelihood that a user viewed sensitive content and whether that access aligns with their established professional role.
System Architecture and Methodology
The proposed auditing framework operates by analyzing query logs and document access patterns to identify the "top-k" most suspicious activities. The system architecture, illustrated in Figure 1, tracks the flow from initial keyword queries to the final identification of anomalous behavior.
The Two-Tiered Scoring System
To distinguish between legitimate professional use and potential privacy violations, the system employs two primary metrics:
- Suspiciousness Score (SScore): This metric estimates the probability that a user actually encountered sensitive information. It is calculated using the IRScore (document relevance) and IRRank (a pagewise exponential decay function). The inclusion of rank acknowledges the "Google effect," where users are significantly less likely to view documents appearing on later pages of search results. If available, the system can also incorporate the time spent on a page to refine this score.
- Anomalousness Score (AScore): This determines if the access was "normal" for a specific user. The system builds a Topic of Interest profile for every user using Topical N-Grams (TNG), which captures technical phrases more effectively than standard word-based models like LDA. Using a k-Nearest Neighbors (k-NN) approach, the system compares a user’s access to their "peers" (others who accessed the same document). If a user’s profile is a distant outlier compared to this neighborhood, the access is flagged as anomalous.
Figure 2 demonstrates how these scores interact. For instance, a nurse (Lucy) might have a high SScore due to a relevant query, but if the access aligns with her role, the AScore remains "No." Conversely, a doctor (Carol) accessing a report outside her department (Gynecology vs. Oncology) might trigger an anomalous flag even with a lower SScore.
Evaluation and Key Insights
Due to the privacy constraints of real-world medical logs, the authors evaluated the system using a surrogate dataset derived from DBLP (3.66 million computer science publications) and WikiCFP. By treating authors as "users" and paper abstracts as "sensitive documents," they simulated anomalous access by identifying instances where researchers in one field (e.g., Databases) accessed papers in unrelated fields (e.g., Music Retrieval).
Key Findings
- TNG Superiority: Topical N-Grams significantly outperformed Latent Dirichlet Allocation (LDA) in distinguishing between normal and anomalous profiles, largely because TNG can capture multi-word technical concepts like "natural language processing."
- Optimal Neighborhoods: Empirical testing identified $k=6$ as the optimal number of neighbors for the k-NN algorithm, providing the best balance across various outlier densities.
- SScore Filtering: The evaluation revealed that true privacy violations tend to cluster around specific SScore thresholds (approximately 0.2). This allows auditors to ignore low-SScore queries, significantly reducing the manual workload required for investigation.
Conclusion and Future Work
The system provides a vital safety net for organizations that must balance data accessibility with privacy. By focusing on the "neighborhood" of users who access a document, the system can distinguish between legitimate cross-departmental collaboration and unauthorized snooping.
Future research may address the "cold start" problem for new users and incorporate more complex, hierarchical organizational structures into the AScore calculation. While the surrogate academic data provided a strong proof of concept, further validation within actual clinical workflows remains a priority for future development.
Summary updated