Speaker
Description
Malware detection using graph neural networks (GNNs) has achieved strong predictive performance; however, these models remain largely opaque and difficult to interpret in security-critical settings. In this work, we propose REx-GNN (Rule-Enhanced Explainable Graph Neural Network), a neuro-symbolic framework that bridges neural predictions and symbolic rules to enable interpretable malware detection. Our approach first extracts candidate rules from a baseline relational GNN through post-hoc pattern mining, and then selects a compact and high-coverage rule subset using greedy and beam search strategies. The resulting set of rules are subsequently injected into REx-GNN training via rule-consistency and explanation-guided losses, promoting alignment between model predictions and symbolic patterns. We conduct a comprehensive empirical evaluation on a heterogeneous ontology-based dynamic malware dataset, examining the trade-off between explanation coverage and false positive rates. Results demonstrate that a small set of optimized rules can explain a large proportion of malware predictions while maintaining low benign trigger rates. Furthermore, we show that beam-optimized rule sets outperform greedy and top-k selection, achieving a superior balance between coverage and noise, and leading to improved model alignment when used in rule-guided training. These findings highlight the effectiveness of integrating rule extraction, optimization, and rule-guided learning as a unified framework for interpretable and trustworthy graph-based malware detection.