Proj CDeepFuzz Paper Reading: CAGFuzz: Coverage-Guided Adversarial Generative Fuzzing Testing of Deep Learning Systems

Abstract

背景：

1. Q: 现有的方法没有考虑到small perturbations的影响，或者这些perturbation只能限定在某个特定的模型上使用，在其他模型上则本身就不符合example要求（可能是指标签就会改变？）
2. 现有的方法多使用浅层的feature constraints来判断生成的adversarial example和原本example的距离，忽略了如object category和scene semantics等的高阶语义

本文：CAGFuzz(Coverage-guided Adversarial Generative Fuzzing)
Task: Generate adversarial examples for DNN models
Method:

1. train an Adversarial Case Generator(AEG) based on general datasets

• Q: AEG特点：只考虑数据特征，忽略低层次的通用能力(only consideres the data characteristics and avoids low generalization ability)

2. 从original and adversarial examples中抽取高级特征(deep feature), 并使用cosine相似度来确保adversarial example和original example的语义相似度
3. use the adversarial examples to retrain the model

实验：
数据集： MNIST(LeNet-1, LeNet-4, LeNet-5), CIFAR-10(VGG-16, VGG-19, Resnet-20), ImageNet(VGG-16, VGG-19, Resnet-50)
Competitors: FGSM, DeepHunter, DeepXplore
效果：在neuron coverage rate, hidden errors 和accuracy上都有所提升

Github: https://github.com/QXL4515/CAGFuzz