Robust Federated-Learning-Based Classifier for Smart Grid Power Quality Disturbances

Area of Science:

• Electrical Engineering
• Computer Science
• Cybersecurity

Background:

• Smart grids require advanced methods for detecting Power Quality Disturbances (PQDs).
• Increased PQDs due to renewable energy and nonlinear loads necessitate robust detection.
• Data privacy concerns limit the use of centralized deep learning for PQD classification.

Purpose of the Study:

• To develop and evaluate Federated Learning (FL) based classifiers for PQD detection.
• To assess the vulnerability of FL models to data poisoning attacks in PQD classification.
• To implement and validate a defense mechanism against poisoning attacks in FL for PQDs.

Main Methods:

• Developed FL-based classifiers for PQD detection and compared them to centralized models.
• Emulated five data poisoning attack scenarios to evaluate model robustness.
• Implemented a detection mechanism to identify and isolate malicious client updates.

Main Results:

• FL models showed a slight performance decrease (97% to 96% accuracy) compared to centralized models.
• Data poisoning attacks significantly reduced classifier accuracy.
• The implemented defense mechanism effectively mitigated performance degradation from poisoned updates.

Conclusions:

• FL offers a privacy-preserving approach for PQD classification in smart grids.
• FL models are vulnerable to data poisoning attacks, impacting classification accuracy.
• The proposed defense mechanism enhances the robustness of FL-based PQD classifiers against adversarial attacks.