Loading...

Evaluating Anomaly Detection Models in Network Time Series by Introducing a New Metric

Alavi, Mostafa | 2025

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57932 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Behroozi, Hamid; Rohban, Mohammad Hossein
  7. Abstract:
  8. With the expansion and increasing complexity of network infrastructures, anomaly detection in network time series has become a key challenge in the fields of system security and reliability. However, existing research in this area faces fundamental issues such as bad-labeled datasets, inappropriate evaluation metrics, and overly complex deep learning methods. Additionally, the lack of sufficient attention to specific scenarios, such as identifying issues related to servers, clusters, and network nodes, raises questions about the effectiveness of current solutions. This study aims to provide a comprehensive approach for anomaly detection in net-work time series, with a special focus on analyzing and improving existing evaluation metrics. First, challenges related to data quality and improper labeling are investigated, and methods for improvement—such as data cleaning, semi-supervised relabeling, and deep data mining—are introduced. Next, through a thorough evaluation of common anomaly detection metrics, their strengths and weaknesses are identified, and new metrics are proposed based on the operational needs of networks and cloud computing services. These new metrics are developed to enhance accuracy in tracking abnormal behaviors, reduce false alarms, and facilitate the detection of server, cluster, and node failures. To demonstrate the effectiveness of the proposed approach, deep models such as Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), and Autoen-coders are utilized. The training and evaluation processes are conducted using real-world and synthetic datasets. The experimental results, in addition to a significant improvement in anomaly detection performance, show the impact of the new metrics in better interpreting results and optimizing network resource management. Finally, recommendations for future research are provided, focusing on adding hybrid solutions and optimizing scalability and efficiency for large-scale systems.
  9. Keywords:
  10. Anomaly Detection ; Evaluation Metrics ; Deep Neural Networks ; Fault Management ; Time-Series Neural Network ; Bad-Labeled Datasets ; Server and Cluster

 Digital Object List

 Bookmark

No TOC