Question about the ood datapoints

[688spiderbin]

Hi, thank for your great work and contributions to point out that kernel shap may be not robust. I have a little question about where can I found the perturbation method source code done by kernel shap？
I notice you put the real xtrain ('background_distribution' here) to the KernelExplainer. To my knowledge, this is all the real data point and then you feed the RF classifier to class ood/iid, my question is where the perturbation datapoint come from?
background_distribution = shap.kmeans(xtrain, 10)
adv_kerenel_explainer = shap.KernelExplainer(adv_shap.predict, background_distribution)
I hope my question is make sense to you !
Thanks !