Efficient Privacy-Preserving and Attack-Detection Solutions for Networked Automation Systems

Abstract

In contemporary cyber-physical infrastructures, networked automation systems play a central role in coordinating the sensing, communication, and control tasks to ensure efficiency, reliability, and safety. However, when untrusted adversaries, such as cloud-based controllers, have access to raw sensor measurements, ensuring security and privacy becomes a critical issue, as these raw sensor measurements contain sensitive information that can be used to launch cyber-physical attacks on the networked systems. To address these concerns, encrypted control and attack-detection methods have been proposed for control and automation systems in the literature. However, both approaches have inherent limitations. Encrypted control solutions can protect raw data confidentiality, but they introduce high communication and computational burdens due to costly cryptographic operations. Attack-detection methods can be used to safeguard control systems against malicious attacks. However, the existing attack-detection methods often ignore the underlying graph structure of the networked automation systems. Although graph-based approaches, such as Graph Attention Networks (GATs), can capture the interaction patterns in automation graphs, their quadratic attention complexity leads to significant latency in large-scale systems, which is detrimental for safety-critical applications.

The objective of this thesis is to address the limitations of the existing encrypted control and attack detection solutions. Throughout the thesis, we use heating, ventilation, and air-conditioning (HVAC) automation as the primary case study to highlight the privacy, efficiency, and security challenges of networked automation systems. In HVAC systems, sensor and actuator networks continuously exchange data with controllers. This information exchange enables optimized comfort and energy usage but also raises the risk of cyber-physical attacks and exposes sensitive information, such as the occupancy level of buildings.

We first introduce a fully homomorphic encryption-based cloud control framework that preserves the privacy of sensitive occupancy information. Sensor measurements are encrypted prior to cloud transmission, ensuring that sensitive information cannot be explored. To reduce the operational cost of encrypted control, we design an optimal event-triggered control policy that transmits data only when necessary, significantly reducing both computation and communication costs. To further protect against information leakage, we propose randomized triggering strategies that obscure timing patterns. Simulation results for a HVAC control application using the TRNSYS platform show that the proposed encrypted control framework achieves effective temperature and CO₂ regulation while reducing encrypted computation and communication costs by at least 60%.

The model-based HVAC control approaches rely on approximating dynamics and are resource-intensive, leading to increasing communication and computation costs. To mitigate this, we remove reliance on the knowledge of explicit building dynamics by designing an encrypted fully model-free event-triggered controller. This controller learns to regulate temperature and CO₂ directly from encrypted sensor data, invoking transmissions only when necessary. Without approximating dynamics, our model-free design reduces the communication by 64% and computation by 75%, outperforming state-of-the-art encrypted model-based methods.

Networked automation systems are vulnerable to cyber-physical attacks such as sensor spoofing and actuator lockouts. We propose an attack detection method for networked automation systems, in which an Event-Triggering Unit (ETU) that locally flags anomalies on the plant side and selectively encrypts and sends data to a cloud-based classifier system. The cloud module combines a GAT, which captures the spatial correlations among HVAC components, with a Long Short-Term Memory (LSTM) network, which models the temporal sequences of encrypted states. Evaluated on diverse HVAC system attack scenarios, the combined GAT-LSTM detector achieves 98.8% accuracy, which significantly outperforms GAT-only (94.2%) and LSTM-only (91.5%) baselines. At the same time, the proposed GAT-LSTM method reduces the data transmission to 15%.

GATs effectively detect attacks by learning to weight interactions across the communication and physical interaction graphs, but their quadratic attention complexity results in high inference latency for large-scale systems. To overcome this limitation, we propose the Gated Hybrid Head Attention (GHHA) mechanism, a novel attention mechanism that uses a learnable gate to combine sparse local attention with global linear attention, reducing per-node complexity to near-linear without sacrificing accuracy. We extend this design to GraphGHHA for graph-structured automation data and evaluate it under diverse building attack scenarios. Our results demonstrate up to an eight-fold reduction in detection latency while maintaining attack detection accuracy. By jointly improving speed and robustness, GraphGHHA enables real-time alarm generation for networked automation systems.

Collectively, the contributions on homomorphic encryption, model-free control, event-triggered communication, and accelerated graph attention address the key challenges of privacy, efficiency, and security in networked automation systems. The results confirm the feasibility of achieving real-time, privacy-preserving, and secure control of networked systems, laying a foundation for an integrated framework that enables privacy-preserving, secure, and efficient networked automation and control.

Date of Award	12 Jan 2026
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Ehsan NEKOUEI (Supervisor)