Template Attacks using Classification Algorithms

Template attacks constitute a powerful side-channel attack technique that is shown to be efficient in breaking many secure cryptographic implementations. This type of attack consists of two phases, the profiling of the device (template building) and the template matching phase. Template matching can create a significant overhead in the performance of the attack, due to the large amount of data that needs to be processed for every possible value. On the other hand, machine learning techniques are developed to provide efficient pattern recognition and feature extraction algorithms, mainly used in artificial intelligence. In this work, we combine techniques from machine learning with template attacks, in order to improve the efficiency of template attacks focusing on the template matching phase. More precisely, we compare three classification algorithms on a template data set built during the execution of a regular scalar multiplication algorithm (double-and-add-always) of mbedTLS (formerly PolarSSL). As a result, we are able to retrieve scalar bits with 20 templates per bit.