二进制神经网络的漏洞研究

项目来源

香(略)资(略)金(略)C(略)

项目主持人

黄(略)

项目受资助机构

T(略)U(略)e(略)t(略)f(略)n(略)o(略)

项目编号

1(略)6(略)

立项年度

2(略)

立项时间

未(略)

研究期限

未(略) (略)

项目级别

省(略)

受资助金额

5(略)7(略)0(略)

学科

C(略)u(略)g(略)i(略)e(略)f(略)a(略)n(略)c(略)l(略)

学科代码

未(略)

基金类别

G(略)r(略)R(略)a(略) (略)d

关键词

未(略)

参与者

P(略) (略)i(略)L(略)

参与机构

未(略)

项目标书摘要:深度(略)攻击大约从5年前开(略)关注它,并且仍然是(略)注点。2019年1(略)文章标题为“为什麽(略)愚弄”的回声表明神(略)脆弱性是一个具有挑(略)题需要一种可靠且有(略)壮性,例如最近开发(略)下限。同时,人工智(略)展是卸载神经网络的(略)练,来自边缘(例如(略)如云端)。这提出了(略)资源限制的边缘设备(略)。为此,轻巧但功能(略)网络已经发展成为传(略)耗同类产品。虽然二(略)全精度相当的高输出(略)络,其对抗攻击的脆(略)研究。      (略)先研究二进制神经的(略)i)为下层提出理论(略)限;ii)调整最小(略)神经网络;和iii(略)找到上层最小对手的(略)的下限(步骤i)直(略)i)和上限受到攻击(略)出现了新的挑战—潜(略)小摄动下限和上限。(略)不存在l_p-范数(略)最新方法计算出的区(略)们的初步研究中,认(略)可忽略下限和基於攻(略)不一致降低了这些最(略)用性。      (略)是双重的,即从理论(略)用漏洞量化,以及设(略)论界限的框架多层感(略)和递归神经网络的方(略)制定了四个主要目标(略)使最新的基於验证的(略)网络;      (略)基於攻击的方法全面(略)制神经网络的脆弱性(略)调查每层神经元激活(略)过缩小每一层的间隙(略)   4)将源代码(略)

Applicati(略): Vulnera(略)y of Bina(略)etworksAb(略)ulnerabil(略) neural n(略)adversari(略)has attra(略)tention b(略)ne learni(略)y since a(略)s ago and(略)important(略)r safety-(略)plication(略)t Nature (略)Oct 2019t(略)eep-learn(略) so easy (略)oes that (略)orks are (略)ersaries (略)ulnerabil(略)as a chal(略)ic.Addres(略)ssuerequi(略)ble and e(略) to evalu(略)ustness o(略),such ast(略) develope(略)lower bou(略)minimum a(略)perturbat(略)le,the la(略)pment in (略)ntelligen(略)yment is (略)e neural (略)putation,(略)nference (略) the trai(略)ntral(e.g(略)the edge((略)evices).T(略)stringent(略)tson the (略)ion due t(略) resource(略)d edge de(略)is end,li(略)erful mod(略)neural ne(略) evolved (略)ze and lo(略)nterparts(略)ional neu(略)s.Althoug(略)ural netw(略)pting in (略)eir equal(略)put accur(略)ble to fu(略)nnetworks(略)erability(略)rial atta(略)n even le(略).To this (略)oject aim(略)nvestigat(略)rability (略)euralnetw(略)ee steps:(略) a theore(略)eneral qu(略) for the (略)of minimu(略);ii)adapt(略) perturba(略)tion sche(略)ry neural(略)nd iii)at(略)ary neura(略)to find an(略) of the m(略)rsary.How(略)paring th(略)lower bou(略) the dire(略)on of the(略)rturbatio(略)nd the up(略)ggested b(略)tep iii),(略)ges arise(略)ially big(略)nthe cert(略)bound and(略)d of the (略)turbation(略)ere are w(略)ute a saf(略)where no`(略)rsaries e(略)ion compu(略)ting stat(略)t methods(略)too small(略)ul.In our(略)y study,t(略)enerally (略)ble gaps (略) certified(略) and the (略)d upper b(略)nconsiste(略)quantifica(略)hes the p(略) of these(略)he-art ro(略)aluation (略)Consequen(略)ls of thi(略)re twofol(略) derive a(略)land gene(略)bility qu(略) for bina(略)etworks,a(略)o designa(略)for tight(略)heoretica(略) state-of(略)wer-bound(略)for multi(略)ptron net(略)lutional (略)orks,and (略)neural ne(略)ccomplish(略)s,four ma(略)es are in(略) adapt th(略)rification(略)erability(略) to binar(略)t-works;2(略)ensively (略)ulnerabil(略)entional (略)neural ne(略) both ver(略)sed and a(略) methods;(略)igate the(略) of neuro(略)ns in eac(略) tighten (略)ounds by (略) gap of e(略))To relea(略)odes and (略)into the (略)in.The PI(略)re recogn(略) machine (略)mmunity a(略)n investi(略)loping,an(略) network (略)ty quanti(略)o-rithms.(略) the PI a(略)e begun t(略)is quanti(略)o theemer(略)f energy-(略)chine lea(略)s such as(略)ralnetwor(略) been pro(略)insights (略) from the(略)es ofmode(略).Research(略) this pro(略)enerate s(略)heoretica(略)cal value(略)AI comput(略)odel port(略) reliabil(略)ritical i(略)

  • 排序方式:
  • 0
  • /
    (略)
  • 排序方式:
  • 0
  • /