Cyber-Physical Attack Detection in Additive Manufacturing: An LSTM-Based Autoencoder Approach Using Reconstruction Error Analysis from Side-Channel Monitoring

In order to identify cyber-physical threats in additive manufacturing systems, this study suggests a sophisticated technique that uses data from side-channel monitoring. Strong attack detection capabilities are guaranteed by the method's integration of many important approaches for pre-processing, analyzing, and classifying time-series data. In order to make the continuous time-series data more understandable and efficient to analyze, a pre-processing method based on defined Window Sliding (FWS) is used to divide the data into manageable windows of defined size. After that, we use Discrete Wavelet Transform (DWT) to get all the features out of each window, collecting important data from all the different frequency bands. Using Particle Swarm Optimization (PSO) to sift through the DWT coefficients and extract the most useful features improves classification performance by making sure that only the most informative characteristics are used. In order to train a deep learning (DL) approach that can identify anomalies via reconstruction errors, the optimal feature set is used. This technique is an LSTM-A autoencoder. We show that the suggested approach can differentiate between normal and attack cases with a 99% accuracy rate by applying it to the problem of side-channel attack detection in additive manufacturing. To protect cyber-physical systems from advanced cyberattacks, this method offers a scalable and adaptable solution while simultaneously increasing detection accuracy.