On the adversarial robustness of DNNs based on error correcting output codes
share
Adversarial examples represent a great security threat for deep learning systems, pushing researchers to develop suitable defense mechanisms. The use of networks adopting error-correcting output codes (ECOC) has recently been proposed to deal with white-box attacks. In this paper, we carry out an in-depth investigation of the security achieved by the ECOC approach. In contrast to previous findings, our analysis reveals that, when the attack in the white-box framework is carried out properly, the ECOC scheme can be attacked by introducing a rather small perturbation. We do so by considering both the popular adversarial attack proposed by Carlini and Wagner (C W) and a new variant of C W attack specifically designed for multi-label classification architectures, like the ECOC-based structure. Experimental results regarding different classification tasks demonstrate that ECOC networks can be successfully attacked by both the original C W attack and the new attack.
READ FULL TEXT