Deep learning has made detecting software vulnerabilities faster and more accurate. However, these models can still be misled by adversarial attacks. Such attacks cause the models to make incorrect predictions, which threatens system security. This study presents a new framework to address these problems. The framework combines Shapley Additive Explanations (SHAP), Support Vector Machines (SVM), and a Bidirectional Long Short-Term Memory (BiLSTM) network with an attention mechanism. Each component helps identify and rank important features, allowing the model to recognize patterns linked to vulnerabilities. In addition, Foolbox generates adversarial examples to test the model’s resilience. This approach strengthens both the accuracy and reliability of vulnerability detection models, building a stronger defense against adversarial attacks.