METHODOLOGY OF GENERATING SYNTHETIC DATA FOR INTELLIGENT ANALYSIS SYSTEMS IN THE PROBLEM OF MALWARE DETECTION

The aim of the work is to develop a methodology for generating synthetic data for malware detection systems. The research method is synthetic data generation using natural language processing methods (T5 transformer and large language model Mistral 7b), originally designed to work with text problems. The result obtained: large datasets are required to expand the range of normal and abnormal behavior. Collecting real data requires a large amount of resources. In this work, one real dataset was collected and 3 synthetic datasets were generated (T5, Mistral 7b, T5 + Mistral 7b). Statistical analysis of the data shows that in most cases the combined dataset (T5 + Mistral 7b) achieves the best results, which is confirmed by a practical experiment. Synthetic data obtained using T5 and Mistral 7b separately are not enough to be used as a training set (a strong decrease in the F1-measure from 0,98 to 0,83 is observed). Using the combined data leads to a result close to the real data (0,98 and 0,97). The scientific novelty consists in determining the possibility of using only synthetic data to train deep learning models, which allows expanding the scope of normal and abnormal behavior in anomaly detection systems.