Proj CDeepFuzz Paper Reading: Testing Deep Neural Networks

Abstract

本文:DeepCover
Github: https://github.com/TrustAI/DeepCover
Task: propose 4 novel test criteria to test DNNs
Method: inspired by MC/DC coverage criteria

  1. Sign-Sign Coverage(Main): the sign change of a condition feature ψk,independently affects the sign of the decision feature ψk+1,j of the next layer
  2. Value-Sign Coverage:
  3. Sign-Value Coverage: captures the significant change of a decision feature’s value that complements the sign change case
  4. Value-Value Coverage: there is no sign change for a condition feature, but the decision feature’s value is changed significantly

SubTask: adversarial testing: symbolic method, gradient-based heuristic search

实验：
数据集：MNIST, CIFAR-10, ImageNet
效果：

  1. 通过这些criteria得到的test samples能够诱发DNN的undesired behavior
  2. 在finding bugs与计算代价之间达到了平衡
posted @ 2023-09-08 02:14  雪溯  阅读(9)  评论(0)  编辑  收藏  举报