A Deep Learning Approach for Malware Detection in IoT Binaries Using Spatial and Temporal Patterns

The proliferation of malware in the Internet of Things (IoT) environment poses significant challenges to IoT security due to the heterogeneity and resource constraints of IoT devices. Traditional malware detection methods, which rely heavily on individual features, static analysis, and raw byte sequences, suffer from performance limitations and are not effective against evolving threats. The proposed work introduces a novel deep learning-based malware detection model that integrates Convolutional Neural Networks (CNNs) and Gated Recurrent Units (GRUs) to learn spatial and temporal representations from binary features. CNN extracts spatial patterns from static binary representations, while GRU extracts sequential dependencies in dynamic binary features, enabling the detection of complex malware behaviors. To further enhance detection efficiency, a feature selection mechanism is incorporated to identify the most relevant spatial–temporal features, reducing training time while maintaining high detection accuracy. The proposed approach effectively combines static and dynamic feature representations to train a robust classifier capable of detecting sophisticated malware patterns. Experimental evaluations on an IoT malware dataset demonstrate the efficacy of the proposed model, achieving an average detection accuracy of 99.33%, significantly outperforming traditional methods. The results also show the model's robustness against obfuscation techniques, with a substantial reduction in the false positive rate (FPR).