Models and algorithms for protecting intrusion detection systems from attacks on machine learning components

Today, one of the means of protecting network infrastructure from cyberattacks is intrusion detection systems. Digitalization requires the use of tools that can cope not only with known types of attacks, but also with previously undescribed ones. Machine learning can be used to protect against such threats. The paper presents models and algorithms for protecting against evasion attacks on machine learning components of intrusion detection systems. The novelty is that for the first time, a simulation of the use of a protection subsystem based on long-short-term memory autoencoders during a fast gradient sign attack was carried out. The methodology consists in simulating adversarial attacks with an assessment of the effectiveness of protection using classical metrics: accuracy, recall, F-measure. The results of the study showed the effectiveness of the proposed subsystem for protecting machine learning components of intrusion detection systems from evasion attacks. The detection indicators were restored almost to their original values.