This project is concerned with finding new methods to improve the test approaches used to validate the certifiability of Deep Neural Networks (DNN) specificities.
Metamorphic testing is a testing technique consisting of using semantic preserving relations in order to generate pseudo-oracle one can leverage to test systems for which actual oracles are hard to get, just as it is the case for DNN. While several techniques applied to DNN already exist, they merely use relations as a proxy for generating new data, and don’t consider the relations as general property of the model. The goal of this project is to study such behavior, in particular how to select and or generate the relations, and how to ensure they are valid with regard to the model one wishes to test them on.
Team
PDF - HOMRS:HighOrderMetamorphic Relations Selector for Deep Neural Networks