人工智能培训

搜索

人工智能教程:PEERNET:利用同伴智慧对抗对手攻击(PeerNets: Exploiting Peer Wisdom Against Adversarial

[复制链接]
~Limi多多~ 发表于 2018-6-4 08:55:37 | 显示全部楼层 |阅读模式
~Limi多多~ 2018-6-4 08:55:37 903 0 显示全部楼层
人工智能教程:PEERNET:利用同伴智慧对抗对手攻击(PeerNets: Exploiting Peer Wisdom Against Adversarial Attacks)深度学习系统已经在我们生活的许多方面变得无处不在。不幸的是,已经表明,这样的系统容易受到对抗性攻击,使得它们容易发生潜在的非法使用。设计对抗攻击强大的深度神经网络是使这种系统更安全并可用于更广泛的应用(例如自动驾驶)的基本步骤,但更重要的是,需要设计新颖的,更先进的基于新计算参数而不是边缘建立的体系结构建立在现有的基础上。在本文中,我们介绍PeerNets,一种新的卷积网络族,将经典的欧几里得卷积与图卷积交替使用,以利用同位素样本的图形信息。这导致了模型中的一种非局部向前传播形式,其中潜在特征以图形引起的全局结构为条件,相比于传统的构造方法,其强度比各种白色和黑匣子敌对攻击强3倍准确度几乎没有下降。
Deep learning systems have become ubiquitous in many aspects of our lives.Unfortunately, it has been shown that such systems are vulnerable toadversarial attacks, making them prone to potential unlawful uses.Designingdeep neural networks that are robust to adversarial attacks is a fundamentalstep in making such systems safer and deployable in a broader variety ofapplications (eg autonomous driving), but more importantly is a necessarystep to design novel and more advanced architectures built on new computationalparadigms rather than marginallybuilding on the existing ones.In this paperwe introduce PeerNets, a novel family of convolutional networks alternatingclassical Euclidean convolutions with graph convolutions to harness informationfrom a graph of peer samples.This results in a form of non-local forwardpropagation in the model, where latent features are conditioned on the globalstructure induced by the graph, that is up to 3 times more robust to a varietyof white- and black-box adversarial attacks compared to conventionalarchitectures withalmost no drop in accuracy.人工智能教程:PEERNET:利用同伴智慧对抗对手攻击(PeerNets: Exploiting Peer Wisdom Against Adversarial Attacks) FPunkVJnPek7PZbJ.jpg
URL地址:https://arxiv.org/abs/1806.00088     ----pdf下载地址:http://arxiv.org/pdf/1806.00088    ----人工智能教程:PEERNET:利用同伴智慧对抗对手攻击(PeerNets: Exploiting Peer Wisdom Against Adversarial Attacks)
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

~Limi多多~当前离线
新手上路

查看:903 | 回复:0

快速回复 返回顶部 返回列表