.png){kind=icon}
.png){kind=icon}
Return
BCI-Fuzz: Bug-triggering Code Innovated to Fuzz Deep Learning Libraries
.png){kind=icon}
PRE
AI
DOI：10.1016/j.jss.2026.112822.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
A considerable number of bugs have been discovered in deep learning (DL) libraries. The resulting crashes or inconsistencies have caused widespread impact on DL systems. The bug-triggering code contains factors that are beneficial for fuzzing, but these code has not been widely utilized in fuzzing DL libraries. Existing methods generate input parameters for APIs of typical code instances through random mutation, resulting in limited code coverage. Simultaneously, extracting API usage specifications from instances means acquiring only partial parameter types of the APIs and covering only APIs present in the instances, which leads to inadequate sufficiency and coverage in API testing. This paper proposes a novel DL library testing method called BCI-Fuzz. Test cases are generated by leveraging bug-triggering code snippets from both structural and semantic perspectives, expanding coverage of unique code behaviors. Meanwhile, parameter constraints extracted can guide the generation of API parameter values, ensuring that multiple parameter types are tested. Inferring similarity between APIs and replacing them in the code allows for covering APIs that are lacking input code invokes. Experiments show that BCI-Fuzz outperforms existing methods, demonstrating practical bug detection abilities. It can obtain guidance information to trigger defects in existing versions from the latest bug-triggering code snippets.
Keywords:
BCI-Fuzz
Deep Learning Libraries
Bug Detection
Fuzzing
API Testing
Journal
IF:
4.1
Papers:
5.4K
Citations:
8.4K
Organization
No organization information available