Fine-tuned transformer models with explainability for Android malware detection

The fast growth of Android apps has raised concerns about safety and privacy, showing the need for simple, reliable, and efficient ways to detect malware. Traditional deep learning models often struggle to capture the long-range patterns to detect complex threats. Transformer based models improve contextual understanding but are still too costly for mobile devices. This study examines six fine tuned Transformer models DistilBERT, TinyBERT, MobileBERT, ALBERT, and CodeBERT along with a new hybrid model called AstraBERT, using a large dataset that contains more than 100,000 APKs from the Koodous platform. By using semantic features like API calls, permissions, and declared actions, AstraBERT achieved the best performance with 95.9% accuracy, 92.1% F1-score, and 97.7% AUC. Combining contextual embeddings with statistical features and using attention visualization through LIME, AstraBERT offers good results. The results show AstraBERT&s;s strength as a scalable, explainable, and high-accuracy system for Android malware detection.