Graph Collection Anonymization via Structural Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for graph privacy anonymization are limited to single graphs and struggle with anonymizing entire collections of graphs, as they rely on proximity notions that are difficult to define and compute, especially in structured data, where structural similarity can manifest as isomorphism, a computationally challenging problem.
Innovation Solution
The approach involves clustering similar graphs based on structural properties to create cluster super-structures that represent aggregate properties, generating synthetic data using probabilistic sampling from these clusters, which retains the aggregate properties of the underlying data for privacy-preserving applications.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing anonymization techniques are applied to single graphs, then privacy protection is achieved for individual nodes, but the ability to anonymize entire collections of graphs is lost
Solution Approach 1:
The patent merges multiple individual graph anonymization techniques into a unified framework that processes collections of graphs. By combining clustering algorithms with existing anonymization methods, the system achieves both privacy protection for individual nodes and anonymization of entire graph collections, resolving the contradiction between reliability of privacy protection and adaptability to different data scales.
Solution Approach 2:
The invention creates a universal anonymization framework that can handle both single graphs and collections of graphs. The system maintains compatibility with existing single-graph anonymization techniques while extending functionality to process multiple graphs through clustering, achieving multi-functionality that addresses the versatility limitation.
2Measurement precision
If structural similarity is used for clustering graphs, then meaningful anonymization clusters are created, but computational complexity increases due to isomorphism checking
Solution Approach 1:
The patent segments the graph comparison process into multiple stages: first computing simplified structural features (degree distributions, path lengths, subgraph patterns) for initial clustering, then applying more rigorous isomorphism checking only within smaller clusters. This segmentation reduces overall computational complexity while maintaining measurement precision for structural similarity.
Solution Approach 2:
The invention changes the parameters used for graph comparison by introducing multiple levels of structural feature extraction. Instead of directly comparing full graph structures, the system transforms graphs into feature vectors based on structural properties, enabling efficient clustering with acceptable precision without the full computational burden of complete isomorphism checking.
3Reliability
If synthetic pseudo-data is generated to preserve aggregate properties, then privacy is maintained, but loss of individual record detail occurs
Solution Approach 1:
The patent applies local quality by generating synthetic data with different levels of detail for different purposes. Within clusters, the system preserves more individual record characteristics, while between clusters, it maintains aggregate properties. This localized approach to data fidelity allows privacy preservation through aggregation while retaining individual record detail where needed for specific analytics.
Data Source
AI summary
Mechanisms are provided for anonymizing data comprising a plurality of graph data sets. The mechanisms receive input data comprising a plurality of graph data sets. Each graph data set comprises data for generating a separate graph from graphs associated with other graph data sets. The mechanisms perform clustering on the graph data sets to generate a plurality of clusters. At least one cluster of the plurality of clusters comprises a plurality of graph data sets. Other clusters in the plurality of clusters comprise one or more graph data sets. The mechanisms also determine, for each cluster in the plurality of clusters, aggregate properties of the cluster. Moreover, the mechanisms generate, for each cluster in the plurality of clusters, pseudo-synthetic data representing the cluster, from the determined aggregate properties of the clusters.