Transformer-aided guard: A hybrid spatio-temporal framework for confidence-aware intrusion detection in cyber-physical systems

In this age of ubiquitous networks, Cyber-Physical Systems (CPSs) have evolved into an essentail part of critical infrastructures like smart grids and Internet of Things (IoT) ecosystems. The interconnectedness of cyber and physical components, however, makes CPSs more vulnerability to sophisticated and stealthy cyber intrusions which are typically untraceable by conventional Intrusion Detection Systems (IDS). Even though Graph Neural Networks (GNNs) are adopted with the aim of strengthening the performance of IDS, there are several challenges noted with available models including limited accuracy. Also, their capacity to capture the dynamic and structured dependencies within CPS environments is insufficient. Within the background, the present study proposes the Transformer-Aided Guard (TAG) approach. This is a novel hybrid spatio-temporal framework, designed for confidence-aware and precise intrusion detection. Through the Swin Transformer and spatial relationship learning using a Spatio-Temporal Graph Neural Network (STGNN), the network leverages temporal sequence modelling. Spatial and temporal embeddings are integrated by a dual-stage infusion mechanism, making it possible to attain more complex learning. Furthermore, a post-hoc Bayesian inference layer based on Monte Carlo Dropout (MC-Dropout) is incorporated to measure prediction uncertainty and enhance decision reliability. Extensive evaluations performed on the benchmark datasets revealed effectiveness of the framework as a promising average accuracy of 98% is achieved. Applicability of the framework is also validated through its detection latency across diverse CPS environments, whereas the ablation study conducted validate the contribution of each architectural component. The findings testify TAG as a secure, scalable, and reliable intrusion detection framework for the next-generation CPS applications.