Enterprise AI Analysis

A deep learning-based IoT malware detection approach for electric vehicle charging stations

This paper introduces a deep learning-based IoT malware detection approach tailored for EV charging stations, with a focus on improving accuracy across diverse CPU architectures. The approach comprises four models. The first model, global structural feature analysis, transforms malware binaries into grayscale images, from which a CNN extracts spatial patterns to generate global structural feature vectors. The second model, Pcode statistical feature analysis, leverages the Ghidra decompiler to unify opcodes from various architectures into platform-independent Pcode representations. These are then vectorized using the TF-IDF algorithm, and a FNN is used to learn their statistical distribution. The third model, Pcode semantic feature analysis, removes redundant Pcode entries and applies a LSTM network to model contextual and semantic dependencies, yielding high-level semantic feature vectors. In the multimodal feature analysis model, the three feature types are fused using a multi-head attention mechanism for dynamic weighting, followed by deep interaction modeling through a multi-layer encoder. This integrated approach captures malware's structural, statistical, and semantic features, enabling precise detection in heterogeneous IoT environments. Experiments conducted on the widely used public IoT malware datasets demonstrate that the proposed approach achieves an average improvement of 1.37% in F1 compared to existing mainstream detection approaches, highlighting its superior performance.