[Paper Review] EM-X-DL: Efficient Cross-Device Deep Learning Side-Channel Attack with Noisy EM Signatures

Josef Danial, Debayan Das+4 more|arXiv (Cornell University)|Nov 12, 2020

Cryptographic Implementations and Security25 references4 citations

TL;DR

This paper presents EM-X-DL, a cross-device deep learning side-channel attack on AES-128 that achieves >90% single-trace accuracy using noisy electromagnetic (EM) traces with low signal-to-noise ratio (SNR). By combining intelligent training device selection, LDA-based pre-processing, and trace averaging, the method enables high-accuracy key recovery on unseen devices with minimal traces, even under challenging low-SNR conditions.

ABSTRACT

This work presents a Cross-device Deep-Learning based Electromagnetic (EM-X-DL) side-channel analysis (SCA), achieving >90% single-trace attack accuracy on AES-128, even in the presence of significantly lower signal-to-noise ratio (SNR), compared to the previous works. With an intelligent selection of multiple training devices and proper choice of hyperparameters, the proposed 256-class deep neural network (DNN) can be trained efficiently utilizing pre-processing techniques like PCA, LDA, and FFT on the target encryption engine running on an 8-bit Atmel microcontroller. Finally, an efficient end-to-end SCA leakage detection and attack framework using EM-X-DL demonstrates high confidence of an attacker with <20 averaged EM traces.

Motivation & Objective

• To address the challenge of cross-device portability in deep learning-based side-channel attacks on low-SNR electromagnetic (EM) traces.
• To overcome the significant signal-to-noise ratio (SNR) degradation in EM side-channel signals compared to power traces, especially due to inter-device variations.
• To develop an efficient, end-to-end framework for EM leakage detection and key recovery on unseen target devices with minimal trace collection.
• To minimize training time and number of required training devices through an intelligent device selection algorithm.

Proposed method

• The method employs a 256-class deep neural network (DNN) trained on averaged EM traces collected from multiple training devices to improve signal-to-noise ratio (SNR).
• Pre-processing techniques including PCA, LDA, FFT, and spectrograms are applied to reduce data dimensionality and enhance feature discrimination.
• Linear Discriminant Analysis (LDA) is identified as the most effective pre-processing method, achieving the highest average cross-device accuracy of ~91.5% with minimal training time.
• An algorithm is proposed to intelligently select training devices based on device similarity, reducing the number of required devices by 20–40% compared to random selection.
• An end-to-end EM scanning framework is used to locate the highest leakage point on a new target device by classifying averaged traces across a grid, with confidence measured by prediction frequency ratios.
• The attack framework leverages the trained DNN to predict the secret key with high confidence using fewer than 20 averaged EM traces on unseen devices.

Experimental results

Research questions

• RQ1 Can a deep learning model trained on EM traces from one set of devices achieve high accuracy when attacking a different, unseen device with low SNR?
• RQ2 Which pre-processing technique—PCA, LDA, FFT, or spectrogram—maximizes cross-device attack accuracy while minimizing training time?
• RQ3 How does inter-device variation affect SNR in EM side-channel attacks compared to power-based attacks?
• RQ4 Can an intelligent device selection algorithm reduce the number of required training devices without degrading attack performance?
• RQ5 Can an end-to-end EM scanning strategy detect the optimal measurement location on a new target device with high confidence using the trained DNN?

Key findings

• The proposed EM-X-DL attack achieves an average cross-device key prediction accuracy of ~91.5% using LDA pre-processing, significantly outperforming other methods.
• LDA pre-processing yields the highest average accuracy (91.52%) and the fastest training time, making it the most efficient approach among tested techniques.
• The method requires fewer than 20 averaged EM traces to achieve high-confidence key recovery on unseen devices, demonstrating strong single-trace attack capability.
• Device selection based on maximal dissimilarity to the training set improves cross-device accuracy compared to selecting similar or random devices.
• The end-to-end EM scanning framework successfully identifies the region of highest leakage on a new chip, with the correct key recoverable from the leftmost quadrant with high confidence.
• The SNR of EM traces drops sharply from 19.6 dB (power) to 3.1 dB (EM) when multiple devices are included, highlighting the critical challenge of inter-device variation in EM SCA.