With the increasing integration of smart grids into our energy infrastructure, the vulnerability of these systems to cyberattacks has become a pressing concern. This research compares and contrasts the results of different anomaly detection models when dimensionality reduction techniques are applied to the data. We aim to conceive effective strategies for real-world cyberattack detection on smart grid systems.

For our study, we worked with a large dataset of more than 3 million records from contemporary cyberattacks such as Heartbleed, Web Attack, Infiltration, Botnet, and DDoS.
To analyze this data, we use four anomaly detection models. Moreover, we were interested in analyzing if there is a significant change in the prediction accuracy of these models when the dimensionality of the data is reduced. We hypothesize that simplifying our data speeds up computations and makes machine learning models work better.
To test our hypothesis, we used different reduction techniques, such as UMAP, PCA, and t-SNE, in conjunction with our anomaly detection models.

Our study shows that we obtained better results when the dimensionality is reduced. This reduction decreases the computational cost of training machine learning models, helps in simplifying the models, and reduces overfitting, leading to better generalization of unseen data.
By focusing on the most relevant features and removing noise and irrelevant information, the model can capture essential patterns in the data more effectively.
Our findings could pave the way for more robust and cost-effective security solutions in smart grid systems.