Boosting Adversarial Attacks on Neural Networks with Better Optimizer

<table class="table-group" id="tab4"><tr><td><table class="table"><tr><td class="thead-hr" colspan="5"><hr/></td></tr><tr class="thead"><td class="align_left">Attack</td><td class="align_center">Inc-v3<sub>ens3</sub></td><td class="align_center">Inc-v3<sub>ens4</sub></td><td class="align_center">Incres-v2<sub>ens</sub></td><td class="align_center">Average</td></tr><tr><td class="thead-hr" colspan="5"><hr/></td></tr><tr><td class="align_left">SI-NI-TI-DI</td><td class="align_center">95.4</td><td class="align_center">93.7</td><td class="align_center">89.5</td><td class="align_center">92.9</td></tr><tr><td class="align_left">SI-AI-TI-DI (ours)</td><td class="align_center">96.2</td><td class="align_center">96.0</td><td class="align_center">92.7</td><td class="align_center">95.0</td></tr><tr class="table-tr"><td colspan="5"><hr class="tbody-hr"/></td></tr></table></td></tr><tr class="table-fn"><td><div>The adversarial examples are generated on the ensemble of Inc-v3, Inc-v4, Incres-v2, and Res-101.<br/></div></td></tr></table>

<div>Comparison of SI-AI-TI-DI and SI-NI-TI-DI under the ensemble-model setting.</div>

Security and Communication Networks

Boosting Adversarial Attacks on Neural Networks with Better Optimizer

Table 4