Deep Learning based IoT-Network Intrusion Detection and Classification Systems: A Survey

This paper surveys the different existing internet-of-things (IoT) intrusion detection systems (IDS). The depth quantification reveals that in the past numerous efforts have been made towards IoT IDS; yet, merely applying standalone convolutional neural network (CNN), recurrent neural network (RNN) or auto-encoders (AE) can’t yield generalizable performance. The traditional CNN and RNN models exploit local hierarchical features and lack contextual details that can impact overall learning and prediction results. Moreover, it undergoes numerous challenges like long-term dependency, gradient vanishing, gradient explode, accuracy degradation etc. Though, the improved deep models like Bi-directional LSTM, Bidirectional GRUs, stacked auto-encoder with ensemble learning can achieve better performance. Unlike traditional deep models where native Softmax classifiers with the cross-entropy cost function is used for classification, the hybrid deep features can be trained over certain optimally designed ensemble learning classifier(s), which can yield higher accuracy with performance generalizability. Additionally, data resampling prior to the feature extraction can alleviate class-imbalance problem and hence can yield higher accuracy. Similarly, feature selection and normalization methods can help to avoid redundant computation while avoiding over-fitting problems. Thus, the strategic amalgamation of the resampling, feature selection, hybrid deep feature extraction and normalization can achieve more efficient and reliable IDS solution. Despite numerous innovations or efforts in the past, none of the state-of-arts could address a key problem that a typical IoT node is designed either to transmit continuously or act as an event-based sensor.