AG-CSE Malware Shield: Adversarial Generation and Chaotic Stacked Ensemble Approach for Deep Learning-Based Malware Classification

Malware attacks on Internet of Things (IoT) devices pose a significant threat due to the rapid proliferation of linked devices and inadequate security measures. IoT device attacks have the potential to destroy device functionality, steal private information, or cause significant disruptions. To guarantee the security of IoT devices, malware detection has been developed. Several techniques were employed to effectively detect malware in IoT devices despite certain drawbacks like poor performance, high processing overhead, limited feature learning capability, overfitting, misclassification, and so forth. To overcome these issues, this research study introduces a novel approach to malware detection in IoT devices by integrating several advanced techniques. The Improved Fast Gradient Sign Method (IFGSM) generates adversarial examples that enhance the model's resilience against potential threats. Data pre-processing through Z-score normalization ensures uniform treatment of features, thereby improving accuracy. Furthermore, the Binary Guided Whale-Dipper Throated Optimizer (BGW-DTO) is utilized for effective feature selection, optimizing model performance by reducing complexity and minimizing overfitting. To classify malware efficiently, the study proposes the Tent Chaotic Jellyfish Optimization-based Stacked Ensemble Fused Deep Learning Model (TcJelo-SEfuse-DL), which combines the strengths of MobileNetV2, Bi-GRU, and Bi-LSTM architectures. The proposed model achieves remarkable accuracy rates of 99.53% on the Aposemat IoT dataset and 99.25% on the BoT-IoT dataset, significantly surpassing existing solutions. This framework offers a comprehensive and effective solution for enhancing malware detection in IoT environments, addressing critical security challenges in this domain.