Leveraging Ensemble Learning for Side Channel Analysis on Masked AES

Symmetric cryptographic functions are known to leak secret information through power side channels. Although masking provides an effective mitigation to such leakage, these protection designs are vulnerable to advanced side-channel analysis based on machine learning. In this work, we propose a framework to leverage ensemble learning for power analysis on a masked AES implementation. Before the key recovery, we use machine learning techniques including decision tree, random forest, BP nerual network, and ensemble learning to recover the masks. We also implement data set enhancement through oversampling techniques including random oversampling, synthetic minority over-sampling technique (SMOTE), k-means-SMOTE, Borderline-SMOTE, and SMOTENC. In addition, we explore different ensemble learning methods on key recovery of power side-channel attacks. Experimental results on DPA contest v4 power traces show that ensemble learning methods improve the attack success rate by approximately 11.29% on average over neural network attacks. The results also reveal that data set enhancement can help reduce the variations in attack success rates between different key bytes.

I. Introduction

Cryptographic functions are usually integrated into critical electronic systems including credit cards, automotive electronics, medical devices, and even high-end weapons to protect confidential information and personal privacy from leaking. These cryptographic functions are secure from the perspective of mathematical theory and computational complexity. For example, RSA is secure because of the difficulty of the factorization of large prime numbers, and ECC security is based on the intractability of the elliptic curve discrete logarithm problem. However, computer scientists discovered a new way to recover the private key from the perspective of side effects of power consumption, timing, and electromagnetic radiation when the cryptographic function executes. Attackers can effectively extract secret information such as keys from electronic devices through technologies such as statistical analysis and machine learning.