Detection and Classification of Block Cipher EM Emissions Using Autoencoder–CNN Framework

This paper presents an automated two-stage framework for detecting and classifying electromagnetic emissions produced during the execution of block ciphers on microcontrollers. The proposed method integrates a one-dimensional convolutional autoencoder for unsupervised anomaly detection with a one-dimensional convolutional neural network classifier for algorithm identification. Real EM traces were collected from STM32L476RG and RP2040 microcontrollers executing AES, DES, 3DES, SM4, IDEA, and Blowfish in ECB mode. The autoencoder achieves reliable segmentation of cryptographic activity without external triggers, while the CNN classifies extracted windows with an average accuracy of 93% across seven classes. The method demonstrates strong robustness to noise, generalizes across devices, and enables non-invasive monitoring of cryptographic activity.