Interpretable CAA classification based on incorporating feature channel attention into LSTM

Abstract

The open and broadcast nature of wireless media makes signal transmission among wireless media prone to different types of channel access attacks (CAA), mainly in Medium Access Control (MAC) layer, ranging from constant jamming to protocol manipulation attacks. CAAs can allow an adversary to greatly degrade overall transmission bandwidth or fully hinder legitimate users from access medium. Therefore, it is critical to timely detect and classify CAAs. A few efforts have been made through applying deep neural networks (DNN) for CAA detection. But they still suffer from low accuracy and poor interpretability. In this backdrop, this paper puts forward an interpretable CAA classification DNN model based on feature channel attention (FCA), named FCA-LSTM. After introducing 11 types of CAAs through state transition model, we detail the design of FCA-LSTM, which incorporates three modules, i.e., FCA module, Long Short-Term Memory (LSTM) module, and Grad-CAM module for promoting classification accuracy while reducing the number of parameters. A series of experiments is conducted to compare FCA-LSTM against four benchmarks, including ResNet50, conventional neural network (CNN), Transformer, and LSTM. Results show that FCA-LSTM performs better than four benchmarks in general. Furthermore, the number of parameters and inference time of FCA-LSTM are both much smaller than traditional LSTM. At last, Grad-CAM is utilized to visualize FCA-LSTM’s concern areas of an input sample. This visualization process sheds light on crucial aspects of model’s decision-making process, further fortifying its interpretability and overall reliability.