Proj CDeepFuzz Paper Reading: Fuzzing Deep-Learning Libraries via Automated Relational API Inference
Abstract
背景：使用API间sharing similar input parameters and inputs的关系能更有效测试DL Libraries
本文: DeepREL
Github: https://github.com/ise-uiuc/DeepREL
Task: inferring relational APIs and fuzzing DL Libraries
Bug Types: Crash, inconsistency
API Match Verifier: 我们将 FreeFuzz 数据库中源 API 的最多 100 次有效调用提供给源 API 和目标 API 作为验证输入，并检查它们是否具有一致的行为(Q: 不是每次执行？）
Step:
- automatically infers potential API relations based on API Syntactic/semantic information
- synthesizes concrete test programs for invoking relational APIs
- validates the inferred relational APIs via representational test inputs
- performs fuzzing to find potential inconsistencies
实验:
对象：TensorFlow, PyTorch
Competitor: FreeFuzz
效果：
- +157% more APIs than FreeFuzz
- detected 162 bugs, 106 confirmed, detected 13.5% high-priority bugs for the entire PyTorch issue-tracking system, detected 14 documentation bugs
