AI-driven cybersecurity framework for anomaly detection in power systems.

Area of Science:

• Cybersecurity
• Artificial Intelligence
• Power Systems Engineering

Background:

• Smart grids face increasing cyber threats like False Data Injection Attacks (FDIA), Denial-of-Service (DoS), and Man-in-the-Middle (MiTM) attacks.
• Traditional security methods are insufficient due to a lack of contextual awareness and real-time adaptability.
• The integration of IoT and automation in smart grids escalates the sophistication and frequency of these threats.

Purpose of the Study:

• To introduce an AI-driven cybersecurity framework for high-accuracy anomaly detection in power systems.
• To fuse cyber and physical datasets for enhanced threat identification.
• To improve the resilience and explainability of smart grid security.

Main Methods:

• Utilized Long Short-Term Memory (LSTM) networks and Random Forest classifiers for anomaly detection.
• Employed SHapley Additive exPlanations (SHAP) for model interpretability.
• Implemented Fast Gradient Sign Method (FGSM)-based adversarial training to enhance robustness against adversarial attacks.

Main Results:

• Achieved 99.798% accuracy in binary classification and 98.1919% accuracy in multi-class classification.
• Improved adversarial accuracy from 95.15% to 99.39% through adversarial training.
• Demonstrated practical viability with model inference completed in 2.16 seconds on a Xilinx PYNQ-Z2 edge device.

Conclusions:

• The proposed AI framework offers high efficacy, explainability, and operational resilience for smart grid cybersecurity.
• The fusion of cyber and physical data significantly enhances anomaly detection capabilities.
• The framework is well-suited for real-time deployment in smart grid environments to ensure secure and reliable power system operation.