Privacy preserving anomaly detection based on local density estimation.

Area of Science:

• Computer Science
• Data Science
• Biomedical Informatics

Background:

• Anomaly detection is crucial across various fields like finance and biomedicine.
• Existing algorithms often suffer from high time complexity and fail to address data privacy concerns.
• Efficient and privacy-preserving anomaly detection remains a significant challenge.

Purpose of the Study:

• To propose a fast anomaly detection algorithm based on local density estimation (LDEM).
• To develop an efficient, privacy-preserving scheme (PPLDEM) for multi-party anomaly detection using homomorphic encryption.
• To reduce communication and calculation costs compared to existing privacy-preserving methods.

Main Methods:

• Introduced LDEM, a novel algorithm utilizing a fast local density estimator.
• Developed PPLDEM, integrating LDEM with homomorphic encryption for secure multi-party computation.
• Estimated local density of instances via the average density across all features, using a defined mapping function for feature density estimation.

Main Results:

• PPLDEM demonstrated lower communication and calculation costs than existing privacy-preserving schemes.
• Security analysis confirmed that PPLDEM does not leak participants' private information.
• Experimental results validated PPLDEM's effectiveness and efficiency in anomaly detection, including activity recognition in clinical settings for older adults.

Conclusions:

• The proposed LDEM algorithm provides a computationally efficient approach to anomaly detection.
• PPLDEM offers a secure and efficient solution for privacy-preserving anomaly detection in multi-party scenarios.
• The PPLDEM scheme is effective for real-world applications, such as analyzing wearable sensor data for health monitoring.