Privacy-preserving outlier detection through random nonlinear data distortion

Area of Science:

• Data Mining
• Privacy-Preserving Techniques
• Anomaly Detection

Background:

• Data owners need to share sensitive data for analysis without compromising privacy.
• Existing methods primarily use linear data perturbations.
• A need exists for more robust privacy-preserving data mining techniques.

Purpose of the Study:

• To explore nonlinear data distortion for privacy-preserving anomaly detection.
• To develop bounds on accuracy and quantify privacy for nonlinear transformations.
• To enable user control over the privacy-utility trade-off.

Main Methods:

• Utilizing nonlinear random data transformations.
• Developing theoretical bounds on expected accuracy.
• Quantifying privacy using standard definitions.
• Applying transformations to linear and sigmoid models.

Main Results:

• Nonlinear distortion proves effective for privacy-preserving anomaly detection.
• The approach allows adjustable privacy levels through nonlinearity.
• Demonstrated effectiveness on real-life datasets.
• Analysis shows specific cases are distance preserving.

Conclusions:

• Nonlinear data distortion offers a flexible and effective method for privacy-preserving anomaly detection.
• The trade-off between privacy and data utility can be managed by adjusting nonlinearity.
• This approach enhances the security of sensitive data analysis.