MC-LRNN: A logic-based neural network for multi-class software vulnerability prediction

Software vulnerabilities are a major threat to information systems. Detecting them early and accurately is critical. Software metrics are commonly used in vulnerability prediction, but choosing the most relevant features remains a major challenge. In this paper, we present Multi-Class Logic Rules Neural Network (MC-LRNN), a novel model that combines logic-based reasoning with neural networks for software vulnerability prediction. MC-LRNN uses a Top-Down Hill-Climbing Greedy Algorithm to extract first-order logic rules from software metrics, forming an interpretable reasoning layer that guides the learning process. The dataset is divided into a Logic Rule Dataset for rule generation and a Learning Dataset for model training and evaluation.

We evaluate MC-LRNN on three benchmark datasets — Juliet, SARD, and REVEAL — under both binary and multi-class classification settings. The results show that MC-LRNN consistently outperforms traditional baselines, handles class imbalance effectively, and generalizes well across projects. Its design provides both interpretability and strong generalization capabilities, making it well-suited for real-world vulnerability prediction. Code and datasets are available at: https://github.com/Seansyx123/LRNN.