A GPU memory leakage code defect detection method based on the API calling feature

A GPU memory leakage code defect detection method based on the API calling feature

연구 분야: Safety

학회: CONF-CDS 2021: The 2nd International Conference on Computing and Data Science

The General-Purpose Graphics Processing Unit (GPGPU) programming has been widely used in artificial intelligence and deep learning, and the GPGPU programming framework, represented by the CUDA framework launched by NIVIDIA Corporation, can apply the powerful parallel computing power of Graphics Processing unit (GPU) to non-graphics tasks. The gradually open computing power of GPU also brings related security risks, but the industry is still mainly concerned with how to dig into the potential security risks of GPU rather than the protection of known problems. In this paper, we propose an API calling feature (ACF) based method for detecting memory leaks in GPU codes and programmatically implement a prototype method to detect the risk of memory data residue in GPU codes written in CUDA framework. The prototype detection method is implemented using the Pass module development capability provided by the LLVM compiler project, and the method is tested to have good accuracy an effectiveness, which can provide a basis for subsequent GPU codes' security research.