Towards a Deep Learning Approach for IoT Attack Detection Based on a New Generative Adversarial Network Architecture and Gated Recurrent Unit

As the use of Internet of Things (IoT) devices has increased rapidly in the last few years, a major challenge is the security of these devices. Machine learning models can adapt to complex malware tactics and identify new forms of malware that may not be detected by traditional methods but the big issue most of cybersecurity solutions face in IoT security is that data is private and contains sensitive information so it can not be available online for cybersecurity specialist in order to use it for training Machine Learning (ML) and Deep Learning (DL) models. There is a problem with the lack of available data for use in ML models. Using Generative Adversarial Network (GAN) can be used to produce simulated data that looks like real one, which can be used to protect the privacy of real-world data. This can be particularly useful in situations where the data contains sensitive information, furthermore IoT devices run using less consumption of computational resources of CPU and memory, so training complex models is a big challenge. To this end, we implemented Gated Recurrent Unit (GRU) instead of the Long Short-Term Memory (LSTM) model because GRU architecture has fewer parameters compared to LSTM and requires less consumption of CPU and memory. We used the UNSW-NB15 dataset in our work. Our results show that this architecture requires less time for training the ML models than LSTM-based models, and in terms of static similarity, majority of columns give a high similarity score. We implemented Random Forest (RF) to classify the dataset. Before generating data, we achieved an accuracy, f1-score, and geometric mean of 97.68%, 97.65%, and 97.7%, respectively. After generating samples using our proposed architecture, we merged these synthetic samples with real ones and classify them using RF. The results improved significantly; we achieved an accuracy, f1-score, and geometric mean of 99.36%, 99.35%, and 99.36%, respectively.