Um fluxo de poda para redes neurais dedicadas a ataques a canais laterais baseados em aprendizado profundo

Abstract

Due to the increasing number of services available over the Internet in recent decades, it is becoming increasingly important to provide security for user information transmitted through various communication channels. To achieve this, computational systems rely on the use of encryption as a means to protect user information. Currently, encryption algorithms are embedded in many devices and applications, which enhances the threat of attacks that exploit the physical characteristics of the hardware executing these algorithms. This class of attack is called Side Channel Attacks (SCAs). Many protections against SCAs, referred to as countermeasures, are found in the literature. However, vulnerabilities are discovered in these countermeasures. In this context, Deep Learning (DL) techniques have attracted increasing interest as they are efficient and continuously evolving resources. Nevertheless, the computational cost of applying DL in the SCA scenario is high. Studies report experiments lasting weeks using computational infrastructure. This work aims to reduce the computational effort of SCA based on neural networks by reducing their size through pruning techniques. Additionally, the computational effort of the network reduction process is optimized. Experimental results demonstrate reductions of 40 to 50% in the number of network parameters, as well as reductions of up to 57.17% in training time. Reduced networks can perform attacks using fewer traces than their respective original networks in all cases. Furthermore, reduced networks require fewer training epochs than their original networks, reducing the time needed to carry out attacks. This thesis demonstrates the potential for an increased threat level posed by DL-based SCAs.