The Problem of Accuracy in Systems for Cyberattack Resistance and the Verification of Neural Networks on the Example of Botnet Detecting Problem

The article considers the problem of the accuracy of detection of intrusions in software systems based on deep learning neural networks. An example of a system for detecting botnets, i.e., malicious software, which are the source of potential attacks, including Denial of Service (DDoS), is presented. The system is created as a classification model that detects the behavior of botnets on infected resources. A number of experiments were conducted on the open data set of the Canadian Institute for Cybersecurity. To increase the accuracy of the classification, the method of augmentating the data set using the method of generating examples of adversarial attacks was used. A method for verification of the reliability of a neural network using automatic proof of the robustness property of the model based on SMT solvers is presented. To increase the accuracy of attack detection, a neurosymbolic approach that combines algebraic methods with classification models is also considered.