Hands-On Artificial Intelligence for Cybersecurity

Adversarial attack transferability

A typical feature of adversarial attacks has to do with their transferability.

This feature refers to the possibility that the adversarial examples generated for a given DNN can also be transferred to another DNN, due to the high generalization capacity that characterizes the neural networks, and that constitutes their power (but also their fragility).

Taking advantage of the transferability of adversarial attacks, an attacker is able to create reusable adversarial examples without needing to know the exact parameters of the individual configurations of the neural networks.

It is therefore very likely that a set of adversarial examples developed to successfully deceive a specific DNN for image classification, ...