CGFuzz: A Dynamic Test Case Generation Method for DL Framework Based on Function Coverage

CGFuzz: A Dynamic Test Case Generation Method for DL Framework Based on Function Coverage

연구 분야: Analysis

학회: 2024 IEEE 24th International Conference on Software Quality, Reliability, and Security Companion (QRS-C)

Deep neural networks have gradually become the cornerstone technology to promote progress in various security fields, and deep learning frameworks are the basis for constructing neural networks. Even minor flaws in a deep learning framework as few as a few lines of code can lead to widespread failures in the model, posing a significant threat to the security of the system. Software testing is an effective means to mine the defects of deep learning framework. The defects originating from the underlying functions of the framework's source code can precipitate problems across all APIs that invoke these functions. This study directs its focus towards the examination of low-level function coverage and proposes an efficient API test case generation technique predicated on the guidance of the coverage. We define the number of low-level functions covered by the execution procedure as a test adequacy metric. At the same time, the method for tracing low-level functions correlated with APIs is proposed, so as to calculate the coverage metric. Based on the completely random generation of API test cases, the coverage is used to guide the seed weight update, so as to optimize the test case selection strategy. Besides, the fuzzy rule selection strategy will be adjusted to perform targeted fuzzy mutation on the selected test cases, leading to efficient test cases which cover more low-level functions. Experimental results underscore the efficacy of our method in generating test cases with enhanced coverage effects, thereby introducing novel methods and concepts for the quality assurance of deep learning frameworks.