Intelligent method for supporting decision-making on software security using hybrid models

Abstract

Objective. The research is aimed at developing an intelligent decision support method for software security assessment using a hybrid model based on deep learning and gradient boosting. The aim is to improve classification accuracy, interpretability and adaptability in the face of growing cyber threats. Methods. The proposed method combines deep neural networks for automated feature extraction and gradient boosting for final decision making. A classification module is built based on calculating the probabilities of software belonging to security classes. In addition, a geometric interpretation of the decision space is used with the calculation of the Euclidean distance to the reference classes (safe, unsafe, uncertain). The probabilities are normalized using the softmax function. The model was trained on a labeled dataset and tested using comparative metrics. Results. The developed prototype demonstrated improved performance compared to classical classification approaches. The experiments confirmed higher classification accuracy and clearer separation of security zones in the normalized feature space. The method effectively identifies cases requiring expert analysis and reduces the frequency of false positives. Visualization of the decision space increases the interpretability of the model results. Scientific novelty. We propose a hybrid intelligent method that integrates two modern machine learning approaches – deep neural networks and gradient boosting – into a single architecture for assessing software security. The decision space is formalized through probabilistic thresholds and geometric interpretation. Practical significance. The method can be used in secure software development processes to automatically assess the level of software security. It supports developers and cybersecurity specialists in identifying potentially dangerous modules at the early stages of the software life cycle. The approach can also be integrated into static analysis systems or CI/CD environments to improve security standards.