Proj CDeepFuzz Paper Reading: Deep learning library testing via effective model generation
Abstract
本文: LEMON
Task: Fuzzing DL Libraries
Method:
- specified mutators for DL Libraries to explore different invoking sequences of library code and hard-to-trigger behaviors
- heuristic strategy to guide the fuzzing to amplify the inconsistent degrees between different libraries and to mitigate the impact of potential noise
Bug Type: Crash, Inconsistencies, Performance Bug, NaN Bug
实验：
对象: Tensorflow, Theano, CNTK, MXNet
效果：
- +24 unknown bugs, + 7 confirmed, +1 fixed
- 证明了heuristic strategy有用
