CURATE: Scaling-Up Differentially Private Causal Graph Discovery.

Area of Science:

• Computer Science
• Machine Learning
• Statistics

Background:

• Causal graph discovery (CGD) estimates probabilistic graphical models from data.
• Differential privacy (DP) is crucial for protecting sensitive observational data in CGD.
• Existing DP-CGD methods apply uniform noise, impacting algorithm performance.

Purpose of the Study:

• To introduce CURATE, a differentially private causal graph discovery framework with adaptive privacy budgeting.
• To address the performance degradation caused by uniform noise in sequential DP-CGD processes.
• To improve utility and minimize privacy leakage in DP-CGD.

Main Methods:

• Developed CURATE, a DP-CGD framework featuring adaptive privacy budgeting.
• Implemented adaptive budgeting to minimize error probability in constraint-based CGD.
• Optimized iteration counts in score-based CGD while bounding cumulative privacy leakage.

Main Results:

• CURATE demonstrated superior utility compared to existing DP-CGD algorithms.
• The framework achieved reduced privacy leakage.
• Experimental validation on multiple datasets confirmed CURATE's effectiveness.

Conclusions:

• CURATE offers an effective approach to balancing privacy and utility in causal graph discovery.
• Adaptive privacy budgeting is key to improving DP-CGD performance.
• The framework provides a more accurate and privacy-preserving method for estimating causal relationships.