Explaining SAR ATR Models Predictions with Counterfactual Explanations

In recent years, Deep Learning models have become more and more popular and their performance have been tested on several domains, from medical to military applications. While being particularly competitive, their non-transparency nature raise plentiful questions, in particular in critical applications for which it seems self-evident to have more insight on models. To alleviate this issue, the eXplainable Artificial Intelligence field has emerged in order to propound more understandable models and consequently to make them more trustworthy. Yet, eXplainable Artificial Intelligence for Synthetic Aperture Radar (SAR) images and especially for Automatic Target Recognition (ATR) is still an emerging subject. In this paper, we investigate the potential of counterfactual explanations to explain ATR model decisions. Not only do we detail the method that we used to generate counterfactual explanations, but we also illustrate how we exploit them to provide semi-global explanations. Thanks to the well-known MSTAR dataset, we effectively show that this approach can unveil how the model imagines a class against another and, more generally, improve the comprehension of the ATR model reasoning.