Umbra: A Visual Analysis Approach for Defense Construction Against Inference Attacks on Sensitive Information.

Area of Science:

• Computer Science
• Data Privacy
• Artificial Intelligence

Background:

• Anonymous data collection is crucial for applications like medical diagnosis and recommendations.
• Traditional privacy methods like syntactic anonymity and differential privacy are vulnerable to sophisticated inference attacks.
• Inference attacks leverage background knowledge and AI models to de-anonymize data.

Purpose of the Study:

• To develop a robust defense mechanism against inference attacks on anonymized personal information.
• To provide a visual interface for understanding and customizing privacy preservation models.
• To ensure that data analysis processes meet stringent privacy requirements.

Main Methods:

• A three-stage approach integrating a visual interface.
• Utilizing Bayesian Networks to depict underlying inference behaviors.
• Implementing customized defenses against unknown adversaries.

Main Results:

• The approach effectively visualizes privacy-preserving model processes.
• It enables users to verify the sufficiency of privacy preservation.
• Demonstrated effectiveness through case studies and expert reviews.

Conclusions:

• The proposed visual, three-stage approach offers enhanced protection against inference attacks.
• It improves transparency and user control in privacy preservation.
• This method strengthens the security of sensitive data in analysis.