Evasion Attacks vs. Machine Learning

Web application demo designed by Ambra Demontis and Marco Melis. Powered by SecML-Lib.

This web application demo creates evasion attacks (a.k.a. adversarial examples) against a multiclass SVM classifier for handwritten digit recognition (using MNIST digits). It implements the high-confidence evasion attacks firstly defined in [Biggio et al., ECML-PKDD 2013 ] for two-class classifiers, and then extended to multi-class problems in [Melis et al., ViPAR 2017 ].

Choose the parameters of the evasion attack below:

Select the perturbation dmax