Investigação do uso de aprendizagem de máquina no fluxo de ataques a canais laterais em sistemas criptográficos

Abstract

Given the advancement of technology and the growing globalization of information, the need for computer systems capable of guaranteeing information confidentiality is intensified. In addition to the continuous development of cryptographic algorithms seeking to hide information, cryptoanalysis techniques evolve to exploit computer systems’ vulnerabilities. In this context, there are attack methods that use the system’s physical properties and have become a concern for designers of cryptographic devices in recent decades. This approach, known as attacks on lateral channels, became famous for being efficient and non-invasive. One of the best-known types is the Differential Power Analysis, which, based on statistical analysis, allows establishing a dependency relationship between processed data and the instantaneous power dissipated by the device. Nevertheless, these attacks require a significant number of consumption traces, capturing samples of the power dissipation obtained during the encryption algorithm’s processing to allow a successful attack. Many works in the literature have been using artificial intelligence techniques, more specifically Machine Learning, to exploit such vulnerabilities to create a model of information leakage that is more efficient than the DPA attack. The primary strategy used is to apply learning algorithms to classify consumption traits according to the Hamming Weight consumption model of intermediate data to the cryptographic algorithm, such as SBOXs functions of the DES and AES algorithms. However, these attack strategies do not exploit the internal data on its own, only the consumption estimate caused by the data. In this context, this work investigates machine learning to create an attack strategy that relates internal data to the algorithm and consumption signatures to find an efficient method as an alternative to DPA attacks. The results obtained show that it is possible to reveal confidential data, such as the values of the keys, inputs, and outputs of the SBOXs of the DES algorithm, in some cases reaching correlation coefficients close to 1, error rates close to 0 using a relatively small number of traces.