Computer Science > Software Engineering
[Submitted on 2 Mar 2019 (v1), last revised 15 Jun 2020 (this version, v2)]
Title:DeepGini: Prioritizing Massive Tests to Enhance the Robustness of Deep Neural Networks
View PDFAbstract:Deep neural networks (DNN) have been deployed in many software systems to assist in various classification tasks. In company with the fantastic effectiveness in classification, DNNs could also exhibit incorrect behaviors and result in accidents and losses. Therefore, testing techniques that can detect incorrect DNN behaviors and improve DNN quality are extremely necessary and critical. However, the testing oracle, which defines the correct output for a given input, is often not available in the automated testing. To obtain the oracle information, the testing tasks of DNN-based systems usually require expensive human efforts to label the testing data, which significantly slows down the process of quality assurance.
To mitigate this problem, we propose DeepGini, a test prioritization technique designed based on a statistical perspective of DNN. Such a statistical perspective allows us to reduce the problem of measuring misclassification probability to the problem of measuring set impurity, which allows us to quickly identify possibly-misclassified tests. To evaluate, we conduct an extensive empirical study on popular datasets and prevalent DNN models. The experimental results demonstrate that DeepGini outperforms existing coverage-based techniques in prioritizing tests regarding both effectiveness and efficiency. Meanwhile, we observe that the tests prioritized at the front by DeepGini are more effective in improving the DNN quality in comparison with the coverage-based techniques.
Submission history
From: Yang Feng [view email][v1] Sat, 2 Mar 2019 09:26:58 UTC (1,875 KB)
[v2] Mon, 15 Jun 2020 09:15:19 UTC (1,743 KB)
References & Citations
Bibliographic and Citation Tools
Bibliographic Explorer (What is the Explorer?)
Litmaps (What is Litmaps?)
scite Smart Citations (What are Smart Citations?)
Code, Data and Media Associated with this Article
CatalyzeX Code Finder for Papers (What is CatalyzeX?)
DagsHub (What is DagsHub?)
Gotit.pub (What is GotitPub?)
Papers with Code (What is Papers with Code?)
ScienceCast (What is ScienceCast?)
Demos
Recommenders and Search Tools
Influence Flower (What are Influence Flowers?)
Connected Papers (What is Connected Papers?)
CORE Recommender (What is CORE?)
arXivLabs: experimental projects with community collaborators
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.