Distributed Online One-Class Support Vector Machine for Anomaly Detection Over Networks.

Area of Science:

• Machine Learning
• Data Science
• Network Security

Background:

• Anomaly detection is vital across many domains.
• One-class support vector machine (OCSVM) is a common approach.
• Distributed data collection via wireless sensor networks is prevalent, requiring online processing of sequential data.

Purpose of the Study:

• To formulate a distributed online OCSVM for anomaly detection in networked environments.
• To develop decentralized algorithms that avoid transmitting raw data.
• To enhance efficiency and accuracy in detecting anomalies from streaming data.

Main Methods:

• Formulation of a decentralized cost function for distributed online OCSVM.
• Utilizing a random approximate function to replace the kernel function for data privacy.
• Incorporating a sparse constraint to find an appropriate approximate dimension.
• Minimizing cost functions using stochastic gradient descent to derive two distributed algorithms.

Main Results:

• The proposed distributed algorithms demonstrate effectiveness through theoretical analysis and experiments.
• Experimental results show low misdetection rates and high true positive rates on synthetic and real datasets.
• Algorithms offer competitive anomaly detection performance with reduced running time and CPU memory consumption.

Conclusions:

• The developed distributed online OCSVM algorithms are effective for network anomaly detection.
• These algorithms provide a practical solution for processing streaming data in decentralized systems.
• The methods achieve a favorable balance between detection accuracy, speed, and resource utilization.