Athena is a novel post-training mutation operator for deep learning models. It aims to inject relevant faults into trained models by leveraging Arachne, search based repair technique for neural networks, to find the most impactful weights given a set of inputs/outputs to degrade.

To accompany the mutation operator, a command-line interface is provided to users as a comprehensive tool for running mutation testing operations on deep learning models. This tool provides commands to generate mutants, run a mutation test on a test set, and evaluate a given mutation operator. The tool is designed to be easily extensible to support new test sets, mutation operators, and models.

To get started, see the User Guide.