人工智能培训

搜索

深度学习论文:通过促进集合多样性提高对抗稳健性(Improving Adversarial Robustness via Promoting Ensemble

[复制链接]
thaihack 发表于 2019-1-28 11:24:26 | 显示全部楼层 |阅读模式
thaihack 2019-1-28 11:24:26 76 0 显示全部楼层
深度学习论文:通过促进集合多样性提高对抗稳健性(Improving Adversarial Robustness via Promoting Ensemble Diversity)虽然深度神经网络已经在变量任务上取得了重大进展,但通常通过模型集合得到增强,但现有的高性能模型很容易受到对抗性攻击。许多努力致力于增强各个网络的鲁棒性,然后构建一个简单的集合,例如,通过直接平均输出,这标志着网络之间的交互。本文提出了一种新的方法来探索各个网络之间的相互作用,以提高鲁棒性的前置模型。从技术上讲,我们在对抗性环境中定义了一个新的集合多样性概念,作为个体成员非最大预测的多样性,并提出了一种适应性多样性促进(ADP)正则化因子,以鼓励多样性,从而通过制造对抗性,从而在整体上产生更好的整体性。个别成员之间难以转移的例子。我们的方法具有计算效率,并且与作用于各个网络的防御方法兼容。各种数据集的经验结果证实,我们的方法可以改善对抗性,同时保持正常实例的最新准确性。
Though deep neural networks have achieved significant progress on varioustasks, often enhanced by model ensemble, existing high-performance models canbe vulnerable to adversarial attacks.Many efforts have been devoted toenhancing the robustness of individual networks and then constructing astraightforward ensemble, e.g., by directly averaging the outputs, whichignores the interaction among networks.This paper presents a new method thatexplores the interaction among individual networks to improve robustness forensemble models.Technically, we define a new notion of ensemble diversity inthe adversarial setting as the diversity among non-maximal predictions ofindividual members, and present an adaptive diversity promoting (ADP)regularizer to encourage the diversity, which leads to globally betterrobustness for the ensemble by making adversarialexamples difficult totransfer among individual members.Our method is computationally efficient andcompatible with the defense methods acting on individual networks.Empiricalresults on various datasets verify that our method can improve adversarialrobustness while maintaining state-of-the-art accuracy on normal examples.深度学习论文:通过促进集合多样性提高对抗稳健性(Improving Adversarial Robustness via Promoting Ensemble Diversity) E6sUpEp8osXFpb6o.jpg
URL地址:https://arxiv.org/abs/1901.08846     ----pdf下载地址:https://arxiv.org/pdf/1901.08846    ----深度学习论文:通过促进集合多样性提高对抗稳健性(Improving Adversarial Robustness via Promoting Ensemble Diversity)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

thaihack当前离线
新手上路

查看:76 | 回复:0

快速回复 返回顶部 返回列表