A Hierarchical Subsequence Clustering Method for Tracking Program States in Spectrograms

Abstract

Electromagnetic (EM) side-channel radiation visualized with a spectrogram can be used to classify program states of a computer processor. However, clustering a spectrogram to automatically track program states is difficult due to their often noisy nature. Popular clustering algorithms like K-Means or HDBSCAN fail to adequately cluster spectrogram samples into the variable-length subsequences that define the program states. These algorithms do not account for the time-continuity of spectrogram samples and consequently tend to assign spurious cluster label changes between samples. Here we develop an algorithm, called Hierarchical Subsequence Clustering for Spectrograms, that uses an intuitive approach to explicitly constrain the clustering problem and generate time-continuous clusters. We demonstrate through experiments with simulated program activity as well as with real EM side-channel data measured from a running cellphone that our automated clustering method is faster and yields better clusters in the presence of significant noise.