Addressing Memorization and Aggregation Risks in AI: A Knowledge Graph Approach to Privacy
Jinhui Zuo
Applied Sciences
Problems Identified (5)
AI memorization privacy risk: AI models can memorize specific records, which can expose sensitive data through model access.
Context-dependent privacy risk modeling gap: Current privacy-enhancing technologies often fail to account for context-dependent privacy risks arising from relationships and interactions between data records.
Private data aggregation risk: Relationships, redundancy, and interactions between data records can create high privacy risks through memorization and data aggregation.
AI memorization privacy risk: AI models can memorize specific records, which can expose sensitive data through model access.
Context-dependent privacy risk modeling gap: Current privacy-enhancing technologies often fail to account for context-dependent privacy risks arising from relationships and interactions between data records.
Proposed Solutions (5)
PrivGraph knowledge graph: PrivGraph is a hierarchically structured knowledge graph for modeling and aggregating private information.
Sensitivity Level Factor: The Sensitivity Level Factor quantifies how much an individual’s private information is embedded in the data.
PrivGraph knowledge probing: A PrivGraph-based knowledge probing method is proposed for post-training privacy assessments.
Lifecycle privacy integration: PrivGraph is discussed as part of the AI engineering lifecycle for traceable full-spectrum privacy protection.
PrivGraph knowledge graph: PrivGraph is a hierarchically structured knowledge graph for modeling and aggregating private information.
Results (3)
Subtle private-link learning observed:
Comparable PII detection performance:
Effective private aggregation modeling:
Research Domain
AI privacy and privacy-preserving technologies