Optimizing 2D CNN Architectures for Tabular IoT Intrusion Data: A Comparative Study Using the BoT-IoT 2020 Dataset

The increasing number of Internet-enabled devices has demonstrated the need to have accurate intrusion detection systems (IDSs). To address this, we adapt the structure of two-dimensional convolutional neural networks (2D CNNs). Particularity, we restructure the inputs and tune convolutional/dense layers (kernel sizes and activations) to non-image features. Using BoT-IoT 2020 dataset, we benchmark the accuracy, precision, recall, F1-score, AUC-ROC, and the false-positive rate (FPR) of a batch-size of 32-1024. The adapted 2D CNNs achieve a 99% accuracy and a lower FPR at a batch size $\geq 128$ , which suggests that it is efficient in detecting IoT-based attacks. In addition, we compared the adapted 2D CNN with powerful tabular baselines, LightGBM has lower FPR (4.8%) and higher AUC (94.8%) at detection accuracy (about 94.7%) whereas MLP/TabNet have high recall, precision and high FPR (about 19.26%), highlighting precision recall/FPR trade-offs.