Abstract
Deep Learning (DL) is an effective way to detect botnet attacks. However, the amount of network traffic data and the required memory space are usually large. Therefore, it is almost impossible to use the DL method on memory-restricted IoT devices. In this paper, we reduce the size of the IoT network traffic data feature using the Long Short-Term Short-Term Memory Autoencoder (LAE) codec section. In order to classify network traffic samples correctly, we analyze long-term variables related to low-dimensional feature produced by LAE using Bi-directional Long Short-Term Memory (BLSTM). Comprehensive testing was performed with BoT-IoT databases to confirm the effectiveness of the proposed DL hybrid method. The results show that LAE significantly reduced the memory space required for data storage of large network traffic by 91.89%, and exceeded the standard features of reducing feature by 18.92 -27.03%. Despite the significant reduction in feature size, the deep BLSTM model shows strength against low model equity and over-equilibrium. It also acquires a good ability to adapt to the conditions of binary classification.