Identificação de criptossistemas pós-quânticos por meio de aprendizado de máquina

Abstract

The development of quantum computing poses a threat to contemporary cryptography, given the capability of these computers to break cryptographic systems based on the practical difficulty of factoring large prime numbers and solving the discrete logarithm problem in finite fields. In response, security institutions have developed post-quantum algorithms, which, like their predecessors, are susceptible to cryptanalytic attacks. Ciphertext-only attack is a type of attack used in cryptanalysis and is considered successful if any information from the ciphertext or the key can be deduced. One of the most important tasks in cryptanalysis is the identification of the encryption algorithm used. The literature review reveals that few studies have exclusively analyzed cryptograms originating from asymmetric ciphers, and none of them have specifically evaluated post-quantum cryptosystems. Therefore, in a ciphertext-only scenario, where only encrypted samples are available to the researcher, this dissertation proposes a methodology based on statistical tests from NIST STS and machine learning algorithms such as Support Vector Machine, Random Forest, Naive Bayes, and K-Nearest Neighbor to identify patterns in the cryptograms and encapsulated session keys from post-quantum cryptosystems Frodo, Crystals Kyber, NTRU, and Saber, in ECB and CBC operation modes, for subsequent identification. As contributions of this research, the detection of patterns in post-quantum cryptosystems is highlighted, considering the state of the art, as well as the development of a method to identify ciphers operating in CBC mode. The research conducted a careful analysis of the influence of the cryptogram size on the performance measures of classifiers, contributing to the advancement of knowledge in this specific area. The results obtained revealed the presence of unique patterns in both the cryptograms and session keys generated by the analyzed cryptosystems, demonstrating the existence of "signatures" in these files. The best performance was achieved by the KNN classifier, which in specific scenarios identified the analyzed algorithms with an accuracy reaching up to 100%.