Malware developers employ advanced API obfuscation techniques, such as name randomization, dynamic resolution, call stack manipulation, parameter spoofing, and API chaining, to bypass detection. Existing unimodal analysis systems frequently fail to identify these threats due to their limited scope. To address this, we introduce a novel multi-modal deep learning framework that combines temporal, structural, and parametric analysis for malware detection. This multi‑modal method uses a temporal encoder with self‑attention to detect name obfuscation and API chaining, Graph attention networks analyze call graphs for call stack tampering and dynamic API resolution, while a contrastive learning module identifies anomalous parameter usage. A key novel approach is the dynamic fusion mechanism, which uses attention-based weighting to combine features, enhancing accuracy and interpretability. Additionally, adversarial training ensures robustness against evasion attempts, with theoretical guarantees on performance under variations. Evaluated on 29,505 real‑world malware samples, the proposed multi‑modal framework achieves a 94.2% F1‑score (an 18% improvement over unimodal baselines) and a 98.1% AUC‑ROC. The framework notably maintains 82% robustness against adversarial variations, significantly outperforming conventional LSTM‑based approaches (45%). Beyond detection, the proposed method provides explainable attention maps for forensic analysis and low latency (<1ms/sample), making it suitable for real-time security deployment. These results suggest that multi-modal fusion is critical for next-generation endpoint protection.

With the rapid increase in malware threats, robust classification methods have become essential to protect digital environments. This study conducts a comparative analysis of machine learning and deep learning methods for malware detection. A variety of models are used from both machine learning and deep learning paradigms to determine their effectiveness in distinguishing malware. To further refine the models, several feature selection techniques are applied to reduce the dimensionality of the data and enhance performance. Performance metrics, including accuracy, precision, recall, and F1-score is used to evaluate each model. The findings indicate that while deep learning approaches generally provide higher detection accuracy, feature selection methods contribute significantly to improving machine learning models in terms of performance and computational efficiency. This analysis offers valuable insights into the balance between model complexity and effectiveness, providing practical recommendations for implementing malware classification systems in real-world applications.