Graph-Native Attention Acceleration for Attack Detection in Cyber-Physical Systems

arXiv:2608.23414 · eess.SY, cs.SY · Submitted 2026-08-24 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Graph-Native Attention Acceleration for Attack Detection in Cyber-Physical Systems".

Paper summary: Rosa: So wrapping up the whole thing, this paper by Feng and Nekouei on "Graph-Native Attention Acceleration for Attack Detection in Cyber-Physical Systems" proposes GraphGHHA. It’s a graph-native attention acceleration layer designed to replace standard graph attention layers in attack detection units.

Dev: The core claim is that this layer combines a local sparse attention branch constrained by the system's adjacency matrix and a global linear mixing branch, both combined by a learnable gate, to get both speed and accuracy.

Taro: What this means for the world is that it makes graph-based anomaly detection practical for real time use in cyber physical systems where you need fast responses to attacks on measurements or control logic.

Rosa: It moves us closer to being able to monitor large scale automation domains like power grids or industrial process control with the low latency required for actual safety alarms.

Dev: They showed an eight-fold reduction in inference latency compared to the vanilla graph attention network, and that complexity is near linear with respect to the graph size s, which is a big improvement over O of s squared.

Taro: The paper's limitation, as they state it, is that while it achieves this speedup and accuracy on simulated scenarios like the HVAC system, extending it to truly dynamic graphs or exploring hardware acceleration for even better scalability will be future work.

Rosa: So, the implication is that this architecture offers a way to keep precision when needed while preserving global context under tight latency budgets in these critical systems.

Conclusion: Rosa: So, we're wrapping up our look at GraphGHHA, this paper by Feng and Nekouei called "Graph-Native Attention Acceleration for Attack Detection in Cyber-Physical Systems."

Dev: Yeah, it's about taking standard graph attention layers and making them way faster without losing the actual ability to spot attacks.

Rosa: It boils down to using a combination of local sparse attention and a global linear mixing branch, all controlled by one learnable gate. It’s designed specifically for attack detection in things like power grids or industrial controls where speed is everything.

Dev: The real win here is the speedup, right? They showed up to an eight-fold reduction in inference time compared to the usual graph attention network. That's a big jump for real-time systems.

Rosa: It means we can run these detectors much more frequently, which is crucial when you need immediate alerts if something goes wrong on the line. But it’s not just about speed; they kept the detection accuracy really high, with that mean F1 score sitting at ninety-seven point six percent.

Dev: That accuracy number is important because sometimes people trade speed for precision, but this layer seems to bridge that gap effectively for graph data. It handles the complexity of the network structure better than older methods we've seen.

Rosa: So what does this mean practically? It suggests that graph-based monitoring in complex physical systems isn't just theoretical anymore; it’s something you can deploy in places like water distribution or power grids where low latency is a hard requirement.

Dev: Exactly. But we still need to keep an eye on the caveats, because they mentioned future work focusing on making it work with truly dynamic graphs and getting hardware support for even bigger systems.

Rosa: Right, so while this gives us a really fast and accurate tool right now for many of these monitoring tasks, the next big challenge is scaling it up to handle systems that are constantly changing.

eess.SY, cs.SY

Submitted: 2026-08-24

Updated: 2026-11-18

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 84/100

The gist: GraphGHHA is proposed as a graph-native attention acceleration layer designed to replace standard graph attention layers in attack detection units to reduce inference latency while maintaining high…

Key concepts

Graph Attention Layers (GAL)
These are the standard deep learning components used to refine node embeddings on interaction graphs. They calculate how important neighboring nodes are when processing information for a specific node, which is crucial for understanding system interactions in cyber-physical systems.
GraphGHHA Architecture
This proposed model introduces a new structure that splits attention into two parts: a local sparse branch that only looks at nearby neighbors and a global linear branch that captures system-wide information. A learnable gate decides how to mix these two branches to balance speed and accuracy.
Sparse Attention Branch
This part of the model limits the attention calculation to a fixed, small set of adjacent neighbors for each node. This constraint significantly reduces computational cost compared to standard attention, making the process much faster for large graphs.
Global Linear Mixing Branch
Instead of calculating complex pairwise interactions across all nodes, this branch uses a kernelized reformulation to compute a linear attention form. This allows the model to retain essential system-wide context efficiently without incurring the high computational cost associated with full graph attention.

Terminology

Summary

GraphGHHA is proposed as a graph-native attention acceleration layer designed to replace standard graph attention layers in attack detection units to reduce inference latency while maintaining high detection accuracy

Motivation and Problem Statement

Cyber-physical systems (CPSs) are vulnerable to attacks on measurements, control logic, and equipment operation Conventional deep learning approaches like MLPs or CNNs are not well-suited for attack detection because they ignore the relational structure between nodes in CPS graphs Graph-based learning is suitable because CPS components interact through communication and physical processes, which can be represented as a graph where nodes are components and edges are interactions However, deploying graph-based detectors in real-time is challenging because the per-node computational cost of existing graph attention mechanisms grows quickly with neighborhood size, leading to inference latency that can reach tens to hundreds of milliseconds This latency is undesirable for safety-critical monitoring and alarm systems where timely detection is essential.

Proposed Architecture

GraphGHHA combines two main branches within a learnable gate to achieve acceleration while preserving accuracy The proposed mechanism consists of (i) a graph-local sparse attention branch constrained by the CPS adjacency matrix, and (ii) a global linear mixing branch to retain system-wide information A learnable gate combines these two branches node-wise to adaptively combine them, preserving detection accuracy under strict latency constraints. The initial feature vector of a node is defined as h(0)i ← [hi1, hi2,. p. 2.., hik]⊤ ∈ Rk The Attack Detection Unit (ADU) applies L graph attention layers (GAL) to refine embeddings as H(l+1) = GAL(l+1)(H(l), G). p. 2

Mechanism Details of GraphGHHA

GraphGHHA decomposes each head into a fixed-size sparse neighborhood branch and a global linear branch, which are combined via a learnable gate The local branch performs adjacency-masked sparse attention over a capped set of adjacent neighbors for each node Specifically, the local branch is expressed as Softmax(Q1K⊤√ 1k/2) V1, T1 ⊙ AD(T1) V2 ≈ [Broadk1 (C1) ⊙ (ReLU2 (Q1) ReLU2 (K1)⊤ V1), Broadk2 (C2) ⊙ (T1 ⊙ A(s/n)D(T1)V2)] The global branch utilizes a kernelized reformulation to compute a linear attention form, which is used in the global branch while preserving a graph-masked softmax in its local branch.

Performance and Evaluation

Experimental results demonstrate up to an eight-fold reduction in detection latency compared to the vanilla graph attention network (GAT) while maintaining high detection accuracy The complexity analysis shows that the total complexity of GraphGHHA is O(s), which is near-linear with respect to the graph size s, contrasting with the O(s 2) cost of standard softmax attention In simulation results on an HVAC system, GraphGHHA achieved up to a 8× reduction in inference latency at the largest scale tested (s′ = 1536), reaching 0.146 ms. Furthermore, the method maintains high detection quality across various attack scenarios, with the overall mean F1-score being 97. p. 96%.

Conclusion and Application

GraphGHHA successfully bridges the gap between graph-based anomaly detection and practical deployment in cyber–physical control systems by achieving near-baseline GAT performance under strict real-time constraints The proposed architecture can be broadly applied to other large-scale automation domains such as water distribution, power grid monitoring, and industrial process control where low latency is required. Future work will focus on extending GraphGHHA to dynamic graphs and exploring hardware acceleration for enhanced scalability. p. 9

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. GraphGHHA enables real-time alarm generation for large-scale Cyber-Physical Systems (CPS) monitoring by reducing inference latency by up to eight times compared to vanilla Graph Attention Networks (GAT). This allows for timely detection and response in safety-critical monitoring applications where delays are unacceptable.

  2. The GraphGHHA architecture, combining a graph local sparse attention branch and a global linear mixing branch with a learnable gate, preserves detection accuracy under strict latency constraints. This design ensures that the system retains both localized interaction sensitivity and system-wide context efficiently.

  3. The proposed method achieves near-linear complexity in terms of graph size s, scaling as O(s), which is significantly better than the quadratic O(s 2) cost of standard softmax attention. This scalability makes it viable for monitoring large control systems like HVAC or power grids with many components.

  4. GraphGHHA can be further optimized using mixed-precision arithmetic, such as FP16 or BF16, to reduce memory traffic by half and speed up inference by about 1.5 times while incurring only a small drop in F1-score (0.2 percentage point). This allows for faster deployment on modern hardware without sacrificing detection quality.

  5. The system can perform multi-class attack identification by extending the Attack Detection Unit (ADU) to output predicted labels for specific attack types, such as actuator, sensor, or controller attacks. This capability allows the ADU to support both binary detection and more detailed multi-class identification.

Sources

Related papers