The growing sophistication of malware has demonstrated major shortcomings in conventional signature-based detection schemes, as well as black-box machine-learning schemes, specifically in terms of being transparent, being resistant to adversarial avoidance, and working reliably. The paper introduces a clear and effective malware detection architecture that can be used to overcome these issues in a real-life cybersecurity situation. The proposed approach a joint implementation of an ensemble-based Light Gradient Boosting Machine (LightGBM) classifier with model-agnostic explainability approaches and adversarial defenses strategies into a single detection system. The classifier is trained on the extracting features of Windows Portable Executable (PE) files which consist of the statistic of the byte level and structural features derived based on the EMBER benchmark dataset. To improve interpretability, the framework incorporates Local Interpretable Model-Agnostic Explanations (LIME) and Shapley Additive Explanations (SHAP), providing both instance-level and global insights into model decisions. Robustness against adversarial evasion is evaluated using surrogate-based gradient attacks, including the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD), together with defensive mechanisms such as adversarial training and defensive distillation. Experimental results show that the proposed framework achieves 91.16% accuracy and a ROC-AUC score of 0.974 while maintaining stable and reliable explanations. Under adversarial conditions, the model preserves significantly higher classification performance compared with non-defended baselines. In addition, a deployment-oriented evaluation demonstrates that the end-to-end pipeline supports near real-time detection with acceptable latency in a simulated cloud environment. Rather than proposing a new learning algorithm, this work focuses on the systematic integration and evaluation of explainability, robustness, and deployment feasibility within a single operational framework. The proposed approach provides a practical and trustworthy solution for AI-driven malware detection in modern cybersecurity systems.