MATHEMATICAL MODELS FOR PREDICTING CYBER-ATTACKS BASED ON ARTIFICIAL NEURAL NETWORKS

Abstract

Introduction. Modern conditions of digital economy development put forward new requirements aimed at preserving economic sovereignty. The increased level of cyber-attacks encourages a transition from reactive to proactive protection based on the use of mathematical methods and models. Integrating such models into national and corporate security systems be-comes an investment in economic resilience, protecting GDP and maintaining public trust in dig-ital ecosystems. Materials and methods. This article examines the problem of predicting phishing and DDoS attacks using machine learning methods, in particular neural networks. The author proposes a three-layer perceptron architecture that is trained using real cyberattack data. The paper examines various activation functions, including Relu, Tanh, Sigmoid, ELU, to determine their impact on the model's efficiency. Research results. A comparative assessment of the constructed models was conducted, demonstrating their performance and accuracy in predicting attacks, which allows us to identify the most effective approaches for protecting information systems. The results of the study may be useful for specialists in the field of cybersecurity and development of threat protection systems. Discussion and conclusion. The experiments conducted showed that each of the activation functions has its own advantages and disadvantages, which allows you to choose the most suitable one depending on the specific task. The article contains graphs that com-pare historical data with forecast values, allowing for a clear assessment of the quality of the predictions. Based on the MAE and MSE metrics, the quality of forecasting cyber-attacks was assessed using the activation functions used in the perceptron model, such as Relu, Tanh, Sigmoid, ELU. This approach allows for a quantitative assessment of forecast data, which ensures an ob-jective comparison of the effectiveness of models.