Research on Key Techniques for Reverse Engineering of Deep Learning Models for x86 Executable Files

In recent years, with the continuous development of machine learning and deep learning, their related applications have gradually appeared in our field of vision, showing explosive growth. A deep learning compiler optimizes a deep learning model, compiles it, and generates a binary file. When this binary file is placed into an edge device, the device can perform model inference. However, the generated binary file shares characteristics with ordinary executable files, meaning that important information within the binary file could potentially be leaked through disassembly. This paper leverages this critical characteristic by reverse-engineering the binary library files of the model to extract key parameters and thereby restore the model's framework. Using model library files deployed on edge devices, combined with disassembly, model prediction, dynamic analysis, and other methods, this study successfully recovers the model's operators, topology, and network hyperparameters. Experimental evaluations demonstrate that the framework of the model can be accurately reconstructed, and the substitute model trained using the original dataset achieves accuracy comparable to that of the original model.