Proj CDeepFuzz Paper Reading: DocTer: documentation-guided fuzzing for testing deep learning API functions

Abstract

背景：现有的constraint-extraction tech 对extracting DL-specific input constraints效率太低
本文：DocTer
Github: https://github.com/lin-tan/DocTer
Task: extract DL-specific input constraints from API documentation
Method: extract API parameter from syntactic patterns using dependency parse tree of API descriptions

实验：
对象：TensorFlow, PyTorch, MXNet
效果：

  1. 85.4% precision in extracting input constraints
  2. detect 94 bugs, +63 bugs, +54 confirmed, +1 CVE, +43 inconsistencies in documents, + 39 inconsistencies fixed
  • baseline(without constraints): +59bugs
posted @ 2023-08-06 20:33  雪溯  阅读(10)  评论(0)  编辑  收藏  举报