MTEDS: Multivariant Time Series-Based Encoder-Decoder System for Anomaly Detection.

Area of Science:

• Computer Science
• Cybersecurity
• Data Science

Background:

• Intrusion detection systems (IDS) are crucial for identifying security vulnerabilities in computer networks.
• Network anomalies, deviations from normal operations, can lead to malfunctions, overloads, and intrusions, disrupting services.
• Time series data, being real-valued, requires specific anomaly detection techniques.

Purpose of the Study:

• To propose a novel multi-variant time series-based encoder-decoder system for anomaly detection.
• To address the challenges of identifying anomalies in complex, multi-variable time series data.
• To improve the stability and traceability of anomaly scores while reducing false positives and negatives.

Main Methods:

• Development of a new multi-variant time series-based encoder-decoder architecture.
• Definition of a novel loss function for updating network weights via backpropagation.
• Evaluation of performance using anomaly scores and comparison with existing methods.

Main Results:

• The proposed system demonstrates superior performance in detecting network anomalies.
• Anomaly scores generated by the system are more stable and traceable.
• The system achieves higher precision, particularly at noise levels of 0.2 and above, compared to Multiscaling Convolutional Recurrent Encoder-Decoder, Autoregressive Moving Average, and Long Short Term Medium-Encoder-Decoder.

Conclusions:

• The proposed multi-variant time series encoder-decoder system is effective for network anomaly detection.
• The novel loss function contributes to improved performance and stability.
• This approach offers a promising solution for enhancing network security through advanced anomaly detection.