performance evaluation of deep learning models on diverse iot datasets for intrusion detection

The internet of things (iot) offers transformative potential across sectors like energy, defense, and healthcare, but its limited resources make it vulnerable to cyberattacks, necessitating robust security measures such as intrusion detection systems (ids) to safeguard its infrastructure. this article presents a study that helps intrusion detection systems identify malicious and legitimate communications. to help the system make the best decisions possible, the subcategory of the attacked traffic is also classified. we trained the suggested models to be capable of binary and multi-class classification, targeting common attacks like denial of service (dos), distributed denial of service (ddos), reconnaissance, and information theft directed at iot devices. our methodology makes use of recently published iot datasets, such as botiot, toniot, wustl-iiot-20212021, and ciciot. to assess and contrast the performance of the proposed models on these datasets, we first applied stratified undersampling to convert the original imbalanced datasets into balanced subsets, which were then used for training and evaluation. among the models evaluated, bilstm achieved the highest accuracy of 99.66% and mcc of 0.99759 on the wustl-iiot-2021 dataset. on the botiot dataset, cnn with dual focal loss reached 97.76% accuracy and 0.95536 mcc. for toniot, lstm achieved 97.01% accuracy with an mcc of 0.93643, while on the ciciot dataset, bilstm obtained 96.23% accuracy and 0.96347 mcc. the results show that bilstm and lstm models give higher performance than fnn and cnn models in terms of precision, recall, f1 score, and mcc across all datasets, demonstrating improved performance for temporal iot intrusion detection tasks.