Logo des Repositoriums

Auflistung nach Autor:in "Nebebe, Betelhem"

1 - 2 von 2
  • Konferenzbeitrag
    AI Defenders: Machine learning driven anomaly detection in critical infrastructures
    (INFORMATIK 2024, 2024) Nebebe, Betelhem; Kröckel, Pavlina; Yatagha, Romarick; Edeh, Natasha; Waedt, Karl
    Previous studies have evaluated the suitability of different machine learning (ML) models for anomaly detection in critical infrastructures, which are pivotal due to the potential consequences of disruptions that can lead to safety risks, operational downtime, and financial losses. Ensuring robust anomaly detection for these systems within a company is vital to mitigate risks and maintain continuous operation. In this paper, we utilize a time-series labeled dataset obtained from a hydraulic model simulator (ELVEES simulator) to conduct a comprehensive and comparative analysis of various ML models. The study aims to demonstrate how different models effectively identify and respond to anomalies, underscoring the potential artificial intelligence (AI) driven systems to mitigate attacks. With the chosen approach, we expect to achieve the best performance in detecting two types of anomalies: point anomaly and contextual anomaly.
  • Konferenzbeitrag
    Integrating Dynamic Thresholding in Anomaly Detection on Water Treatment Facilities
    (INFORMATIK 2024, 2024) Yatagha, Romarick; Oeztuerk, Esin; Nebebe, Betelhem; Edeh, Natasha; Mejri, Oumayma
    With the growing complexity of cyber-physical systems (CPS), adaptive and robust monitoring solutions are increasingly crucial for ensuring operational reliability and safety. Anomaly detection is a critical component of monitoring systems, particularly in dynamic environments such as water management systems, where operational regimes can vary significantly over time. Traditional static thresholding techniques, which use a single fixed threshold for the entire monitoring process, are often inadequate due to their inability to adapt to changing data patterns, leading to high rates of false positives and missed detections. This paper explores the limitations of static thresholding and presents a comparative analysis with more adaptive approaches. We first discuss the use of static thresholds for each regime shift, which provides some improvement but still falls short in accommodating gradual or unexpected changes. Subsequently, we introduce a dynamic thresholding method based on the Autoregressive Integrated Moving Average (ARIMA) model. This approach continuously adjusts thresholds in real time, effectively accounting for evolving data patterns and regime shifts. Our evaluation, conducted on synthetic water level data with known anomalies, demonstrates that dynamic thresholding significantly outperforms static methods. Specifically, dynamic thresholding achieves an accuracy of 99%, precision of 78%, recall of 88%, and an F1-score of 82%, highlighting its robustness and adaptability. These results underscore the potential of dynamic thresholding techniques to enhance anomaly detection in complex, variable environments.