Proj CDeepFuzz Paper Reading: Coverage Guided Differential Adversarial Testing of Deep Learning Systems

Abstract

本文：DLFuzz
Task: Fuzz Machine Learning Models using differential testing to detect inconsistency before/after perturbation with minutely mutating
Method: minutely mutating to maximize the neuron coverage and the prediction difference, neuron selection strategies
Pros: Without manual labeling or cross-referencing oracles
实验：
datasets: MNIST(LeNet-1, LeNet-4, LeNet-5), ImageNet(VGG16, CGG19, ResNet-50)
Competitors: DeepXplore
效果：

1. could generate 338.59% more adversarial inputs with 89.82% smaller perturbations
2. averagely obtain 2.86% higher neuron coverage, and save 20.11% time consumption
3. neuron selection performs better than DeepXplore(?)
4. adversarial retraining能提升效果