Exploring the recurrent and sequential security patch data using deep learning approaches

Abstract

The ever-changing nature of vulnerabilities and the intricacy of temporal connections make the classification of security patch data, both sequential and recurrent, a formidable challenge in cybersecurity. The goal of this research is to improve the efficacy and precision of security patch management by optimizing deep learning models to deal with these issues. In order to assess their performance on the PatchDB dataset, four models were used: recurrent neural networks (RNN), long short-term memory (LSTM), gated recurrent unit (GRU), and bidirectional LSTM (Bi-LSTM). Metrics like F1-score, area under the receiver operating characteristic curve (AUC-ROC), recall, accuracy, and precision were used to evaluate performance. When it came to processing sequential data, the GRU model was the most efficient, with the best accuracy (77.39%), recall (65.63%), and AUC-ROC score (0.8127). With a 75.17% accuracy rate and an AUC-ROC score of 0.7752, the RNN model successfully reduced false negatives. With AUC-ROC scores of 0.7792 and 0.8055, respectively, LSTM and Bi-LSTM had better specificity but more false negatives. To improve cybersecurity operations, decrease mitigation time, and automate the classification of security updates, this study presents a methodology. To improve the models' practicality, future efforts will center on increasing datasets and testing them in real-world settings.