Prototype-Based Dynamic Autoencoders for Real-Time Anomaly Detection and Explainability

Real-time detection of anomalies in data streams is a foundation of modern applied analysis in complex systems. It enables experts to design rapid, efficient, reliable, and high-performance decision support systems. Rapid identification of abnormal or unexpected behavior is often insufficient in high-risk contexts; therefore, it is essential to provide explanations for these predictions. Although the use of autoencoder-based algorithms has been shown to be effective in the detection of anomalies, it is important to note that their effectiveness decreases with time due to changes in the underlying data distribution. To solve this drawback, we introduce a solution that incorporates the dynamic updating of autoencoders based on the prototypes computed from their latent representations. We propose two algorithms: Update-Based Prototypes-v1, an algorithm specifically designed to detect anomalies through continual, prototype-informed updates, and ARCANA Prototypes-v1, which is developed to provide reliable explanations for anomalies detected over time. We used the Skoltech Anomaly Benchmark database for our experiments. The Update-Based Prototypes-v1 algorithm achieved a detection rate of 80% with a false alarm rate of 20%, compared to 49% for an implementation without updates. The results confirm that the proposed model offers high performance comparable to state-of-the-art architectures in the same field, even when data characteristics change. Additionally, ARCANA Prototypesv-1 provides precise explanations, which enhances the usefulness of the prototypes for explaining anomalies.