search icon
email icon
rrss gif icons
twitter icon
linkedin icon youtube icon

Deep learning assisted side-channel analysis evaluation, becoming friend of a misunderstood monster

Servio Luis Paguada Isaula


  • DIRECTORS: dr. Igor Armendariz, Prof. dr. Lejla Batina and dr. Ileana Buhan
  • UNIVERSITY: Radboud University Nijmegen


Side channels are unintended interfaces that an electronic device has. A side channel could be used to interpret the state of the electronic device or what the device is processing. For instance, the temperature of the device is an indicator of the device’s state. Then, we can say the temperature is a side channel of the device and the heat is the side channel information. In the past decade, the side channel analysis grew in attention as was discovered that side channels are a source of vulnerabilities for the cryptographic algorithms implemented in the electronic devices. The so-called side-channel attacks exploit those vulnerabilities giving to an attacker the possibility to recover the sensitive information that the device should keep secret. Side-channel attacks have a great impact in the cybersecurity field as they compromise the valuable information of the users and companies.

To address this problem several procedures, tests and countermeasures evaluation have been included as topics of study into the side-channel analysis. Side-channel analysis evaluation study the procedure and methodologies to evaluate the threat of a side channel over a device. 
Several statistical tools and algorithms have been incorporated in side-channel analysis evaluation. In particular, deep learning models have been proved to be one of the best algorithms to perform side-channel evaluation due to its capability to overcome the difficulties in the interpretation of the ground true of the side channel information. 

However, develop a deep learning network to efficiently evaluate a side-channel attack is challenging. There are several aspects to consider when design a deep learning model, from the architecture of it to the value of its hyperparameters. Moreover, a single deep learning model is not aimed to evaluate several cryptographic implementations as they are trained to fit a particular implementation.

Thus, this thesis focuses on applying new process and methodologies, as well as modifications to the deep learning architecture used in side-channel analysis evaluation. The main research topic of this thesis could be divided in two lines: first, it proposes the integration of methodologies and frameworks to improve the training process of the deep learning model. Second, it proposes new ways to design the architecture of the deep learning model. Several tools and analysis are presented including six sigma methodology, early stopping framework, feature reduction and transfer learning approaches.


close overlay