Proj CDeepFuzz Paper Reading: Regression Fuzzing for Deep Learning Systems
Abstract
本文：DRFuzz
Task: find the regression faults between versions of a DL system
Method:
- a diversity-oriented test criterion to explore as many faulty behaviors as possible(⼀种⾯向多样性的测试标准来探索尽可能多的错误⾏为).
- 基于GAN的fidelity保证机制
实验：
数据集：四个回归场景（补充训练、对抗训练、模型修复和模型剪枝）中的四个主题(MNIST LeNet-5, CIFAR10 VGG16, Fashion-MNIST AlexNet, SVHN ResNet18)
竞争对象：DiffChaser, DeepHunter
效果：在检测到的回归错误数量⽅⾯平均提⾼了 1,177% 和 539%。
III. Approach
B. GAN-based Fidelity Assurance
It is also important to ensure the fidelity of fault-triggering test inputs since it is easy for test inputs with low fidelity to fool a DL system but are not concerned by developers due to being out of the scope of the DL system [20]. Therefore, test inputs with high fidelity are more meaningful for regression fuzzing in practice.
C. Heuristic-based Regression Fuzzing Process
奖励函数的设计有两个⽬标，即变异规则在版本之间带来较⼤预测差异的概率和变异规则⽣成⾼保真度测试输⼊的概率
