How to Accurately and Privately Identify Anomalies.

Area of Science:

• Data Science
• Computer Science
• Cybersecurity

Background:

• Anomaly detection is crucial for science, security, and finance.
• Privacy concerns limit the analysis of sensitive datasets.
• Existing methods like differential privacy struggle with accurate anomaly identification.

Purpose of the Study:

• To develop a novel privacy framework, sensitive privacy, for anomaly detection.
• To demonstrate the limitations of differential privacy in this context.
• To propose practical algorithms with strong privacy and utility guarantees.

Main Methods:

• Introduced the concept of sensitive privacy, generalizing differential privacy.
• Developed a compiler to transform differentially private mechanisms into sensitively private ones.
• Proposed specific mechanisms for (β, r)-anomalies.

Main Results:

• Sensitive privacy offers robust privacy and utility for anomaly detection.
• Differential privacy is insufficient for accurate and private anomaly identification.
• The proposed mechanisms achieve high accuracy and provable guarantees.

Conclusions:

• Sensitive privacy is a necessary advancement for private anomaly detection.
• The developed mechanisms provide a practical solution for sensitive data analysis.
• This work enables more effective anomaly detection across various critical domains.