Proj CDeepFuzz Paper Reading: NeuRI: Diversifying DNN Generation via Inductive Rule Inference

Abstract

背景：The correctness of DL systems is crucial for trust in DL applications
本文: NeuRI
BaseTool: FreeFuzz
Github: https://github.com/ise-uiuc/neuri-artifact
Task: Fuzzing DL Libraris and explore more types of operators
Bug Type: 1. Crash(includes unexpected Python exceptions, not include interpreter exceptions) 2. inconsistency between compiler and interpreter 3. Sanitizer error:ASan, UBSan, CSan
Step:

1. collecting valid and invalid API traces from various sources, 就是freefuzz
2. apply inductive program synthesis over traces to infer the constraints
3. hybrid fuzzing

实验：
对象：PyTorch, TensorFlow
Competitor: NNSmith
效果：

1. +51%, 15% branch coverage
2. detected 87 bugs, 64 confirmed, 8 high-priority bugs