Adaptive Bayesian Networks for Online Anomaly Detection

Abstract

This paper proposes a novel approach to online anomaly detection utilizing adaptive Bayesian Networks (ABNs). Traditional Bayesian Networks (BNs) often fail to effectively handle the dynamic nature of real-world data streams, leading to inaccurate anomaly detection. Our approach overcomes this limitation by dynamically updating the conditional probability tables within a BN using a reinforcement learning (RL) algorithm. The RL agent is guided by anomaly signals, allowing the network to continuously learn and adapt to the evolving underlying data distribution. We define a formal framework for this adaptive BN, outlining the key components and their interactions. The core concept involves constructing a BN where node parameters, specifically the conditional probabilities, are continuously refined. This creates a self-adapting system capable of identifying deviations from the learned "normal" behavior. We demonstrate the feasibility and effectiveness of this approach through a theoretical analysis and provide a detailed description of the algorithm. The resulting adaptive system offers improved responsiveness and accuracy compared to static BNs in online anomaly detection scenarios. The system's ability to learn and adjust its internal parameters based on incoming data makes it particularly well-suited for environments where the normal behavior is not stationary. The primary contribution lies in the synergistic combination of Bayesian inference and reinforcement learning, providing a robust and adaptable solution for real-time anomaly detection.