Instruction Level Disassembly through Electromagnetic Side-Chanel: Machine Learning Classification Approach with Reduced Combinatorial Complexity

EM side-channel can be quite effective at instruction level disassembly of the executing program. This leaks IP from Internet of Things (IoT) networks. This may also serve as a benign capability to reverse engineer IoT malware binaries. Power Side Channel instruction level disassembly state-of-the-art is capable of identifying instructions in a 2-3 stage pipeline at 50-200 MHz clock frequency with reasonable accuracy by grouping instructions. EM side-channel works at distance unlike power side-channel. Machine Learning models for instruction identification, Principal Component Analysis (PCA) for feature selection, Gaussian Process Classifiers (GPC), Adaptive Boosting (AB), Quadratic Discriminant Analysis (QDA), Naïve Bayes (NB), Support Vector Machines (SVM) and Convolutional Neural Network (CNN) for instruction classification were developed. Our results of implementation on a 2-stage pipelined architecture demonstrate that the EM side-channel classification approach identifies instructions in flight with 99% accuracy.