NeuroScope: Reverse Engineering Deep Neural Network on Edge Devices using Dynamic Analysis

The usage of Deep Neural Network (DNN) models in edge devices (e.g., IoT devices) has surged. In this usage scenario, the inference phase of the DNN model is executed by a dedicated, compiled piece of code (i.e., a DNN binary). From the security standpoint, the ability to reverse engineer such binaries (i.e., recovering the original, high-level representation of the implemented DNN) enables several applications, such as stealing DNN models, gray/white-box adversarial machine learning attacks and defenses, and backdoor detection. While a few recent works proposed dedicated approaches to reverse engineer DNN binaries, these approaches are fundamentally limited in the type of DNN binaries they support. To address these limitations, in this paper, we propose NEUROSCOPE, a novel data-driven approach based on dynamic analysis and machine learning to reverse engineer DNN binaries. This compiler-independent and code-featurefree approach enables NEUROSCOPE to support a larger variety of DNN binaries across different DNN compilers and hardware platforms, including binaries implementing DNN models using an interpreter-based approach. We demonstrate NEUROSCOPE's capability by using it to reverse engineer DNN binaries unsupported by previous approaches with high accuracy. Moreover, we showcase how NEUROSCOPE can reverse engineer a proprietary DNN binary compiled with a closed-source compiler and enable gray-box adversarial machine learning attacks.