TEST-INT: A Testing Platform for Deep Learning Models

Ozdemir Ö., Demir D.

15th Turkish National Software Engineering Symposium (UYMS), ELECTR NETWORK, 17 - 19 November 2021, pp.155-157, (Full Text) identifier identifier

  • Publication Type: Conference Paper / Full Text
  • Doi Number: 10.1109/uyms54260.2021.9659790
  • Country: ELECTR NETWORK
  • Page Numbers: pp.155-157
  • Middle East Technical University Affiliated: No

Abstract

Deep learning models indicate remarkable performance in a wide variety of tasks especially in computer vision. However, it is often that deep learning models developed only perform well on a specific and limited test data, and fail in real-world applications. In safety-critical applications, it is highly significant deep learning model. The TEST-INT system, which we have produced to solve this problem, is a testing platform that creates new test sets from existing test data with different image transformation methods and adversarial attacks, as well as measures test adequacy and performance of the model, and reports them to the user.