Machine Learning-Based AES Key Recovery via Side-Channel Analysis on the ASCAD Dataset

TL;DR

This paper demonstrates that machine learning models, especially CNNs and RF with feature selection, can effectively exploit electromagnetic side-channel leaks to recover AES keys, highlighting practical vulnerabilities in cryptographic implementations.

Contribution

It introduces a machine learning framework using RF, SVM, CNN, and ResNet models with feature importance analysis for side-channel key recovery on the ASCAD dataset.

Findings

RF with feature selection achieves Rank 0 with fewer traces

CNN reaches Rank 0 efficiently with about 65 traces

ResNets perform best on complex datasets but less so on simple fixed key data

Abstract

Cryptographic algorithms like AES and RSA are widely used and they are mathematically robust and almost unbreakable but its implementation on physical devices often leak information through side channels, such as electromagnetic (EM) emissions, potentially compromising said theoretically secure algorithms. This paper investigates the application of machine learning (ML) techniques and Deep Learning models to exploit such leakage for partial key recovery. We use the public ASCAD `fixed' and `variable' key dataset, containing 700 and 1400 EM traces respectively from an AES-128 implementation on an 8-bit microcontroller. The problem is framed as a 256-class classification task where we target the output of the first-round S-box operation, which is dependent on a single key byte. We evaluate standard classifiers (Random Forest (RF), Support Vector Machine (SVM)), a Convolutional Neural Network(CNN) and a Residual Neural Network(ResNet). We also explore the utility of RF-based feature importance for dimensionality reduction. Crucially, we employ this domain-specific Key Rank metric for evaluation, showing its necessity over standard classification accuracy. Our results show that SVM and RF on full features perform poorly in key ranking. However, RF trained on reduced (top 100) identified via importance analysis achieves Rank 0 (successful key byte recovery) using almost half the attack traces. The implemented CNN also achieves Rank 0 efficiently using approximately 65 attack traces for the fixed-key dataset. The ResNets perform best on large and complex datasets but may not always be the best choice for simple fixed key dataset in terms of efficiency. Thus we conclude that models, particularly CNNs, ResNets and feature-selected RF, coupled with the Key Rank metric, are an effective tool for side-channel key recovery, confirming the practical vulnerability of the cryptographic implementations.