Temporal Modeling of Data Pipeline Anomalies with Spiking Neural Networks

Abstract

This research investigates temporal anomaly detection employing Spiking Neural Networks (SNNs), leveraging biologically inspired, event-driven computation to model complex temporal dependencies. The SWaT dataset, consisting of multivariate time-series data under normal and attack conditions, served as the experimental benchmark.

Preprocessing encompassed Kalman filtering for missing sensor data, forward-fill for binary actuator states, cyclical encoding of timestamps, and independent normalization. The data was divided into training, validation, and test sets with a 60–20–20 split, refraining from shuffling to preserve temporal structure. Multistep windowing facilitated the detection of both point and collective anomalies, with labels generated exclusively for the validation phase, reflecting predefined attack scenarios.

Experimental findings reveal that SNN-based models adeptly capture temporal dynamics and surpass conventional deep learning methods in anomaly detection accuracy and latency. The central objective was to empirically evaluate Maass’s hypothesis regarding the computational equivalence or superiority of spiking neural networks relative to classical neural architectures. These results underscore the promise of neuro morphic computing for real-time anomaly detection in critical infrastructure systems.