Proj CDeepFuzz Paper Reading: Fuzz Testing based Data Augmentation to Improve Robustness of Deep Neural Networks

Abstract

背景：

1. 对微小扰动十分脆弱主要来自于overfitting、limited datasets
2. 由于Specification往往不完备，因此automatic program repair通常要从程序行为中学习

本文：SENSEI, SENSEI-SA
Github: https://github.com/gaoxiang9430/sensei
Task: use fuzzing to augment the training data of DNNs
Method: 将data augmentation program整体视作一个优化任务，用genetic search来生成input data, Q: 试图识别能够skipping augmentation的机会来加速

实验：
datasets: GTSRB(1,2,3,4), FM(1,2,3), CIFAR-10(1,2,3,4), IMDB(1,2), SVHN(1,2)
Competitors: Random augmentation, Worst-of-10, yang2019invariance
效果：

1. improve the robust accuracy by 11.9%-5.5%
2. reduce the average DNN training time by 25% while still improving robust accuracy