arXiv:2608.23414v2 Announce Type: replace Abstract: Cyber-physical systems (CPSs) rely on extensive networks of sensors, controllers, and actuators, making them highly vulnerable to attacks targeting measurements, control logic, and equipment operations. While graph-based attack detectors, especially graph attention models, efficiently localize such attacks by exploiting the communication and physical interaction graphs. The computational cost of edge-wise attention scales poorly with the number of interactions. This bottleneck severely limits the scalability of online monitoring in large or densely connected CPSs. To address this challenge, we introduce GraphGHHA, a graph-native attention acceleration layer designed to reduce the inference cost of graph attention mechanisms in CPS security. GraphGHHA consists of a local attention unit (LAU) and a global attention unit (GAU). The LAU captures the local interactions by restricting the attention computation of each node of a CPS to a limited number of its neighbors, whereas the GAU captures the system-wide interactions. Dynamic gates then weight and combine these local and global representations. We evaluate GraphGHHA on a networked heating, ventilation, and air conditioning (HVAC) cyber-physical system under representative attack scenarios. Based on our results, the inference latency of GraphGHHA is 89.3\% smaller than that of the standard graph attention layers (for an interaction graph size of 1536), while achieving an accuracy of 0.982 in detecting attacks. These results demonstrate that the proposed attention mechanism can significantly reduce the attack detection latency without compromising accuracy.