Proj CDeepFuzz Paper Reading: DeepHunter: A Coverage-Guided Fuzz Testing Framework for Deep Neural Networks
Abstract
本文: DeepHunter
Task: Fuzzing Deep Learning Models
Github: https://github.com/Shimmer93/Deephunter-backup
Method:
- Metamorphic mutation to generate new semantically preserved tests
- use multiple plugable coverage criteria as feedback to guide the test generation
- maintains multiple tests in a batch
- prioritizes the test selection based on active feedback
实验:
datasets: MNIST(LeNet-1, LeNet-4, LeNet-5), CIFAR-10(ResNet-20, VGG-16), ImageNet(MobileNet, ResNet-50)
Feedback criteria:
- Neuron Cov. (NC) The ratio of activated neurons;
- K-multisec. Neu. Cov. (KMNC) The ratio of covered k-multisections of neurons;
- Neuron Bound. Cov. (NBC) The ratio of covered boundary region of neurons;
- Strong Neuron Act. Cov. (SNAC) The ratio of covered hyperactive boundary region;
- Top-k Neu. Cov. (TKNC) The ratio of neurons in top-k hyperactived state;
- Bottom-k Neu. Cov. (BKNC) The ratio of neurons in top-k hypoactived
效果:
- 证明可以增加coverage
- 能生成useful tests
- 能为platform migration准确capture potential defects