Design of Active Learning Framework for Collaborative Anomaly Detection

In this paper, we propose a collaborative anomaly detection system based on the Active Learning framework. In this system, multiple Edge nodes are responsible for monitoring and collecting data from the local users. To reduce the data to be transmitted, we adopt the Principal Component Analysis(PCA) method in the data preprocessing stage on each Edge node. Then each Edge node selects the most valuable samples based on the Uncertainty Sampling strategies. These samples are then uploaded to the Cloud server for anomaly detection based on the machine learning classification schemes. The samples are also used to train the sample selection model on the Cloud. The parameters of the Cloud classifier and sample selection model are then sent back to the Edge nodes to update their classifiers and sample selection models. This system allows us to reduce the cost in sample obtaining, shorten the bandwidth demand and lower the latency. We implement the proposed anomaly detection schemes in a practical system and provide experiment results to demonstrate the performances of the system.