How can edge AI improve CI security in a sustainable way?

What makes securing cyber-physical systems so challenging?

Critical infrastructure has evolved into cyber-physical systems (CPS), where digital and physical components interact dynamically. This transformation enhances operational efficiency, but also creates new cyber risks. Swiftly detecting and localizing anomalies is crucial to preventing major disruptions, yet traditional centralized intrusion detection systems struggle to address the unique demands of CPS. Centralized models often rely on large data centers, consuming significant energy and contributing to carbon emissions. This is where edge computing and the concept of “Green AI” can make a meaningful
impact.

How can edge computing make critical infrastructure more sustainable? How can edge computing make critical infrastructure more sustainable? 

Our research in Resilmesh includes the development and testing of edge AI models for cyber attack detection in critical infrastructure. Our initial experiments used a publicly available critical infrastructure dataset, with the idea of proposing a more sustainable approach that can be applied across different sectors. By leveraging edge computing, we aim to process data closer to its source, reducing reliance on energy-intensive centralised servers while improving cybersecurity.
Our experiment involves splitting the dataset into three distinct zones, each managed by a potential edge device. These edge devices operate independently, training local models and making real-time decisions, which significantly reduces data transmission. This decentralized approach not only decreases latency but also lowers the carbon footprint associated with system operations, aligning with the principles of Green AI by advancing sustainability in AI-driven processes.

Key Advantages: Improved Localization and Efficiency

1. Enhanced Anomaly Detection and Localization: Decentralizing detection allows for more precise identification of threats, as local edge devices can better recognize the specific areas affected by anomalies. This leads to faster responses and targeted interventions, minimizing damage and maintaining the integrity of critical systems.
2. Energy Efficiency and Reduced Carbon Footprint: By processing data locally, edge computing achieved a 75% reduction in carbon emissions during training, with similar detection performance to centralized models. Inference mode showed even greater efficiency, with energy savings and emissions reduced by two orders of magnitude, making it a highly sustainable option for real-time anomaly detection in critical infrastructure.
3. Faster Detection Speeds: While the resulting 1.5% improvement in detection speed may appear minor, it is crucial in cyber-physical systems where even slight delays can escalate incidents. The combination of quick detection and accurate localization enhances the system’s resilience to cyber threats, ensuring minimal disruption to essential services.

What’s next on our green journey? 

Although our research focused on a specific dataset, the principles underlying our decentralised model are broadly applicable to different types of critical infrastructure. The deterministic nature of the physical processes in these systems supports the generalisability of our findings. In the future, we aim to extend our experiments to additional datasets, including recordings from real systems, to explore the capabilities of different hardware modules that fit into edge environments, and to further improve edge-based anomaly detection.