AEG: An Adversarial Attack Algorithm Based on Structured Data and Threshold Constraint

Although machine learning models are very effective in practical applications, many researches have shown that most algorithms are vulnerable to attack with well-designed adversarial examples and adversarial attack algorithms. But the current studies are mainly based on images, which are not suitable for models trained on structured data. In this paper, to expand the research on adversarial examples applicable to structured data, a new adversarial attack algorithm called AEG is proposed with a modified gradient descent method. Real-world structured data and trained load prediction models are applied for attack experiments. Experiments show that the algorithm has achieved significant attack performance while evading the anomaly detection mechanism.