Mitigating Over-Generalization in Anomalous Power Consumption Detection using Adversarial Training

Power consumption anomaly detection systems that use neural networks for prediction tasks are vulnerable to adversarial attacks, leading to unreliable performance and potential adverse effects on the power grid. Certain attack configurations can evade detection or trigger false alarms due to the neural networks’ generalization tendencies, causing adaptation to attack values. This study investigates the effectiveness of adversarial training methods in improving detection performance against such attacks on power consumption data. Leveraging techniques like the Fast Gradient Sign Method (FGSM), Basic Iterative Method (BIM), and Projected Gradient Descent (PGD), we assess the resilience of anomaly detection models. Through empirical experiments, we evaluate detection accuracy, adjustment capabilities, and prediction errors of these adversarially trained models across three datasets. Our results show significant improvements in detection performance, particularly in attack scenarios that normal prediction models would not detect. Additionally, we analyze the models’ adaptability to anomalous data and quantify prediction errors, providing insights into their robustness and limitations. Integrating adversarial training techniques into anomaly detection models for power grids can reduce over-generalization to attack data, enhancing the detection of malicious demand manipulation attacks.