Download PDFOpen PDF in browserEA-NET: A Hybrid and Ensemble Multi-Level Approach For Robust Anomaly Detection10 pages•Published: November 24, 2022AbstractIn the current world, the applications of anomaly detection range from fraud detection to diagnosis in the medical area. Most of the current methodologies are applicable only when a particular dataset pertains to certain assumptions and a distinct domain. Such assumptions require prior knowledge of the dataset. The training development cycle time to find the best single model is time-consuming and challenging. Unsupervised anomaly detection methods do not use the target label for training. However, they result in high false positive rates. In this paper, we address the problem of the ensemble anomaly detection approach that generalizes well across multiple domains. We design a multi- level hybrid approach. At the first level, we train several weak classifiers (weak one class classifiers). Next, we utilize deep learning-based AutoEncoder to reduce the dimension of the dataset. These are the two sets of hybrid features. Next, different one-class classifiers have their strength and limitations. Thus, we propose an adaptive weightage approach that gives the weight to each classifier. Next, this input is passed to the second level. At this level, we have a deep neural network that learns the patterns of the dataset and generates an adaptive dynamic threshold to discriminate the input feature as an anomaly or benign. The major benefit of this approach is the low false-positive rate. The training time is reduced due to the reduction of the input feature dimensions at the first level.Keyphrases: autoencoder, ensemble anomaly detection, one class classifier In: Frederick Harris, Alex Redei and Rui Wu (editors). Proceedings of 31st International Conference on Software Engineering and Data Engineering, vol 88, pages 18-27.
|
