Defogger: A Visual Analysis Approach for Data Exploration of Sensitive Data Protected by Differential Privacy

Area of Science:

• Computer Science
• Data Privacy
• Human-Computer Interaction

Background:

• Differential privacy (DP) is crucial for data security but hinders interactive data exploration due to privacy budget limitations and noisy query results.
• Existing exploration strategies often lack flexibility and can lead to user confusion when dealing with uncertainty inherent in DP systems.

Purpose of the Study:

• To address the challenges of data exploration under differential privacy.
• To propose a novel visual analysis approach for formulating exploration strategies tailored to user intent.
• To enhance the usability and effectiveness of data exploration in privacy-preserving environments.

Main Methods:

• Described exploration scenarios, requirements, and strategies for differentially private data analysis.
• Developed a visual analysis approach integrating a reinforcement learning (RL) model to suggest exploration strategies.
• Incorporated a novel visual design for representing uncertainty in correlation patterns within the prototype system.

Main Results:

• The proposed visual analysis approach, powered by RL, offers diverse strategy suggestions aligned with user exploration intent.
• The integrated uncertainty visualization aids users in understanding and navigating noisy correlation patterns.
• User and case studies demonstrated the approach's effectiveness in developing satisfactory exploration strategies.

Conclusions:

• The visual analysis and RL-based approach effectively supports users in formulating exploration strategies under differential privacy.
• This method enhances the practical application of differential privacy by improving the data exploration experience.
• Future work can further refine the RL model and uncertainty visualization for more complex exploration tasks.