Deep Neural Networks (DNNs) are proprietary assets due to the expertise, confidential data, and high development costs involved in model training. Well-trained DNN models are compiled into DNN binaries to be efficiently executed on various platforms, such as edge devices and cloud infrastructures. Recent research on DNN binary decompilation shows the potential of stealing DNN models via binary reverse engineering techniques. While obfuscation is a well-studied technique to hamper binary reverse engineering, general obfuscation schemes are not designed for this new type of binary and have limitations in concealing information within DNN binaries due to the unique characteristics of DNN binaries.

Intel Software Guard eXtension (SGX) is the cornerstone of Confidential Computing, enabling runtime code and data integrity and confidentiality via enclaves. Unfortunately, memory-unsafe and type-unsafe programming languages, such as C/C++, are commonly used to develop enclave implementations. As a result, a memory corruption or a data race within enclaves could lead to different attacks against the enclaves, such as Return-Of-Programming (ROP) and data leakage, breaking the hardware security guarantee provided by SGX. To automatically identify these issues in existing enclave implementations, in this paper, we propose FuzzSGX, an input and program mutation-based fuzzer for Intel SGX enclave implementations. FuzzSGX provides an enclave fuzzing runtime, FuzzSGX Runtime, a drop-in library for Intel SGX SDK, enabling code coverage and sanitization within enclaves. To explore the host app-enclave boundary, FuzzSGX conducts static analysis and symbolic execution on existing host apps and enclave implementations to generate promising fuzzing programs, fuzzing both ECALLs and OCALLs. We evaluate FuzzSGX using 30 popular SGX applications and enclave implementations and find 93 bugs among these SGX projects, including data races, null pointer dereferences, out-of-bound accesses, division-by-zero, etc. FuzzSGX achieves 3.2x higher code coverage and finds 48.2% more bugs by directly targeting the host appenclave boundary by using program mutations, compared to state-of-the-art fuzzers.