Improving robustness of Transformers for power quality disturbance classification via optimized relevance maps

Power quality disturbances (PQDs) are a critical area of research in the power systems domain, as they can significantly affect the stability of the power grid. Deep learning (DL) models have recently demonstrated outstanding performance when applied to PQD classification. However, they are notoriously susceptible to adversarial attacks, thus posing a tremendous cyber-physical risk to the power grid. While traditional defense methods like adversarial training may increase the robustness of DL-based PQD classifiers, they require substantial computational resources and introduce the accuracy-robustness trade-off. In this work, we propose a new approach to increase the robustness of a family of PQD classifiers against adversarial attacks by increasing the explainability of the contextual relations between different features of the input signal. To this end, we use the Transformer encoder model for PQD classification, relying on its self-attention mechanism to introduce a relevance property to this model. We show that by incorporating the resulting explainable features in the form of relevance maps into the loss function during the fine-tuning stage, we achieve improvement in the robustness and resiliency of the model against adversarial attacks, without compromising accuracy, and without investing additional computational resources required for adversarial training.

M.Sc. student under the supervision of Prof. Yoash Levron.