Differential-Private Data Publishing Through Component Analysis.

Area of Science:

• Computer Science
• Data Privacy
• Machine Learning

Background:

• Data publishing requires balancing user privacy with data utility.
• Existing privacy-preserving data publishing (PPDP) methods often compromise data utility.
• Differential privacy offers a robust privacy guarantee.

Purpose of the Study:

• To propose novel PPDP mechanisms using component analysis to enhance data utility while maintaining differential privacy.
• To evaluate the performance of differential Principal Component Analysis (PCA) and differential Linear Discriminant Analysis (LDA) for PPDP.

Main Methods:

• Developed differential PCA-based PPDP for general data dissemination.
• Developed differential LDA-based PPDP for classification tasks.
• Compared proposed mechanisms against state-of-the-art methods and traditional differential privacy mechanisms (Laplacian, Exponential).

Main Results:

• Differential PCA-based PPDP demonstrated improved utility (smaller error) compared to Laplacian and Exponential mechanisms for the same privacy budget.
• Differential LDA-based PPDP showed effectiveness for classification-oriented data dissemination.
• Both mechanisms outperformed existing methods in specific use cases.

Conclusions:

• Component analysis offers a viable approach for achieving a better trade-off between privacy and utility in data publishing.
• Differential PCA and LDA provide effective PPDP solutions with enhanced data utility.
• The choice between differential PCA and LDA depends on the specific data dissemination objective.