Abstract
Mobile malware detection is difficult despite great accuracy in benign conditions. This study presents a hybrid architecture that detects fraudulent Android apps while resisting adversarial perturbations using Long Short-Term Memory networks, Residual networks, and multi-head attention methods. AE-LSTM-ResNet uses sequential temporal dependencies from LSTM units and hierarchical feature representations from residual connections with attention mechanisms to target crucial malware indications. On the Drebin dataset of 11,120 samples from 179 malware families, the model achieves 97.38% clean accuracy and 94.53% adversarial accuracy, a 4.12% improvement over previous methods. Systematic assessment against FGSM, PGD, and AutoAttack is included. Certified robustness bounds provide realistic security assurances, and extensive tests show improved robustness-accuracy tradeoff. Through thorough adversarial robustness evaluation, our study bridges theoretical detection accuracy and actual deployment security.