Leveraging Trusted Input Framework to Correct and Forecast the Electricity Load in Smart City Zones Against Adversarial Attacks

Recent cybersecurity research is shifting to favor adversarial attacks, which are harder to detect and counteract, as they are not apparent, and they exploit the vulnerability of the deployed machine learning model. With machine learning seeing wider adoption to forecast the electricity load in smart cities to help maintain grid stability and better integrate with renewable energies, the energy sector is at risk. Although several methods exist to mitigate the problem, adversarial training fails against the attack that utilizes a surrogate model. In addition, it cannot differentiate between the original and adversarial data without relying on adversarial detection. Furthermore, existing detections do not correct the data to keep the forecast running. In this paper, the Multivariable Convolutional Encoder implementation in a Trusted Input Framework was proposed, which utilizes hidden input data with a high correlation to act as a trap when creating the surrogate model, measuring the attack intensity, and improve the accuracy to “correct” and “forecast” on the target data. In the experiment, the proposed method not only fooled the surrogate but also achieved a coefficient of determination score of 0.9183 on the most robust Projected Gradient Descent-based adversarial attack, which is 13.933% better than the adversarial training.