Uncertainty estimation based adversarial attack in multi-class classification

Abstract Model uncertainty has gained popularity in machine learning due to the overconfident predictions derived from standard neural networks which are not trustworthy. Recently, Monte-Carlo based adversarial attack (MC-AA) been proposed as a simple estimation method is powerful capturing data points that lie overlapping distribution of decision boundary. MC-AA produces uncertainties by performing back-and-forth perturbations given point towards boundary using idea attacks. Despite its efficacy against other methods, this only examined on binary classification problems. Thus, we present and examine with multi-class tasks. We out limitation multiple classes tackle converting multiclass problem into ‘one-versus-all’ classification. compare recent model methods Cora – graph structured dataset MNIST an image dataset. conducted experiments performed variety deep algorithms perform Consequently, discuss best results LEConv AUC-score 0.889 CNN 0.98 methods.