A Platform for Characterizing Artificial Intelligence Safety and Robustness developped at CEA List


AIMOS (AI Metamorphism Observing Software) is a tool which goal is to provide the means to test metamorphic properties on a dataset for a given AI model. The approach is black-box as the intrinsic characteristics of the AI model is not used in any way, as the model is treated as a black-box oracle with only its inference function.

Its approach improves the usual approaches which only test a model against a perturbation without considering its output. Indeed, a transformation on the input space can also be linked to a transformation on the output space, e.g. a symmetry on the input can mean a symmetry on the outputs as well (a left arrow becoming a right arrow, etc.). AIMOS implements this approach, thus allowing the user to provide more flexible and complete property testing. As the tool is agnostic of the model type used, this allows the approach and framework to compare different types of models or architectures and, in turn, to provide incentives for a choice between one type or another. As such, this permits comparisons between similar models types but with different architectures (e.g. 1 Convolution vs 2 Convolution layers) but also between very different architectures (e.g. Neural Networks vs SVMs vs Decision Trees) in similar settings or also of simply differently trained models.

The aim of this tool is to rapidly and widely test a given AI model on a given dataset. This testing can serve to exhibit inefficiencies of the models, compare them to gauge the most stable ones, etc. Thus, AIMOS can be an important part of an AI development process, providing a framework to facilitate and automate a rapid testing procedure. The metamorphic properties themselves, as part of the testing procedure, should be carefully selected to provide a good criterion of evaluation for the model when tested with AIMOS. For instance, if the problem represented by the AI model does not present any axe of symmetry, it would be of little value to test it against that.