Evolutionary Model Validation—An Adversarial Robustness Perspective

WhenAdversarial robustness building Machine Learning models, either manually or automatically, we need to make sure that they are able to solve the task at hand and generalize, i.e., perform well on unseen data. By properly validating a model and estimating its generalization performance, not only do we get a clearer idea of how it behaves but we might also identify problems (e.g., overfitting) before they lead to significant losses in a production environment. Model validation is usually focused on predictive performancePredictive performance, but with models being applied in safety-critical areas, robustness should also be taken into consideration. In this context, a robust model produces correct outputs even when presented with data that somehow deviates from the one used for training, including adversarial examplesAdversarial examples. These are samples to which small perturbations are added in order to purposely fool the model. There are, however, limited studies on the robustness of models designed by evolution. In this chapter, we address this gap in the literature by performing adversarial attacks and evaluating the models created by two prominent NeuroEvolutionNeuroevolution methods (DENSER and NSGA-Net). The results confirm that, despite achieving competitive results in standard settings where only predictive accuracyPredictive accuracy is analyzed, the evolved models are vulnerable to adversarial examplesAdversarial examples. This highlights the need to also address model validation from an adversarial robustnessAdversarial robustness perspective.