A Lightweight Attention-Enhanced Deep Learning Framework for Malware Detection in IoT: A Comparative Study of Structured and Image-Based Data Representations

The runaway growth in IoT and IIoT devices has revolutionized sectors like healthcare and manufacturing while increasing the attack surface for cyber-attacks. The connectivity explosion highlights the need for effective, IoT-focused intrusion detection and mitigation systems (IDMS). At their core is having access to realistic, heterogeneous datasets that reflect actual network behaviour and attack patterns. Recent data sets cater to these requirements by providing increased protocol support, device heterogeneity, and diverse attack settings, which enhances both federated and centralized IDMS research. In order to overcome existing scalability and interpretability issues, a new lightweight deep learning framework is presented that relies on the use of autoencoders for compressing features and attention mechanisms to bring emphasis to essential patterns in IoT traffic. A compact multi-layer perceptron is used to classify with less computational expense suitable for deployment at the edge. The model is evaluated on both structured and image data formats, showing high accuracy in malware detection with performance sufficient for real-time IoT intrusion detection.