四川大学学报（自然科学版）2026，Vol.63Issue(2)：259-274,16.DOI:10.19907/j.0490-6756.250305
Cuda-Gen:一种基于API知识图覆盖驱动的CUDA模糊测试框架
Cuda-Gen:An API-knowledge-graph coverage-driven fuzzing framework for CUDA libraries
摘要
Abstract
In the AI-driven era,NVIDIA CUDA libraries have become indispensable for accelerating compute-intensive tasks,yet their security assessment remains critically understudied due to closed-source code and unique program.gming paradigms.Existing efforts primarily focus ontarget CUDA compiler vulner-abilities(e.g.,NVCC),but leaving library-level risks largely unexplored.overlook broader library-specific risks.This The paper addresses the challenges of fuzzing CUDA libraries:1)the absence of guidance nar-rows the set of APIs that generated harnesses can reach;and 2)input mutation remains inefficient for LLM-generated harnesses.We propose Cuda-Gen,Aa new tool called Cuda-Gen has been proposed,aimed at un-covering potential vulnerabilities in the CUDA libraries.Cuda-Gen has the ability tocan generate testing har-nesses for various CUDA library functions from scratch,perform efficient parameter mutation,and adapt to the needs of multiple CUDA libraries.First,LLMs are used to extract semantic relationships from CUDA documentation and sample codes,constructing a knowledge graph that prioritizes API interactions and contex-tual dependencies.We introduce anThe API coverage bitmap is proposed to guide the fuzzer to explore under-tested library functions.AdditionallyBesides,we integrate the API knowledge graph is also combined with compiler diagnostics to automatically repair erroneous harnesses,thereby improving compilation success rates.Subsequently,Cuda-Gen employs the LLMs to analyze and decouple parameter dependencies,sepa-rates out the mutable parameters,and performs parameter-isolated mutation on them to enhance mutation effi-ciency.Evaluated across three CUDA releases(12.4,12.7,and 13.0)on eightsix widely adopted libraries(e.g.,cuBLAS,cuFFT),Cuda-Gen achieves on average 2.97×improvements inhigher API coverage and 4.0×improvements insuperior API edge coverage over the baseline tool Fuzz4Allrelative to baseline(Fuzz4all),on average.The experiments uncovered 43 unknown vulnerabilitiesbugs,validated by NVIDIA's security team.关键词
模糊测试/基于大语言模型的知识图谱/CUDA库安全/API覆盖位图Key words
fuzzing/LLM-based knowledge graph/CUDA libraries security/API coverage bitmap分类
信息技术与安全科学引用本文复制引用
宋霁洋,范希明,高心怡,丁雪川,雷琦,方勇..Cuda-Gen:一种基于API知识图覆盖驱动的CUDA模糊测试框架[J].四川大学学报（自然科学版）,2026,63(2):259-274,16.基金项目
军委科技委预先研究项目(CA131B0502T-065) （CA131B0502T-065）
