“Imperceptible,Robust,and Targeted Adversaria lExamples for Automatic Speech Recognition”
背景:
1、對(duì)抗樣本大多用于圖像領(lǐng)域;
2、目前用于音頻的對(duì)抗樣本有兩個(gè)缺點(diǎn):
(1)容易被人類察覺
改進(jìn)方法:頻率掩蔽。通過使用另外一種充當(dāng)“掩蔽器”的信號(hào)對(duì)對(duì)抗性樣本進(jìn)行掩護(hù)
(2) 在空氣中傳播時(shí)不太起作用
改進(jìn)方法:
攻擊原理:
Given an input audio waveform x(輸入音頻)
a target transcription y (目標(biāo)轉(zhuǎn)化結(jié)果)
an automatic speech recognition (ASR) system f(·) (語(yǔ)音識(shí)別系統(tǒng))
a small perturbation δ
objective is to construct an imperceptible and targeted adversarial example x0→ x0 = x + δ 通常通過執(zhí)行梯度下降( gradient descent)來生成對(duì)抗性示例
? Targeted: the classi?er is fooled so that f(x‘) = y and f(x) != y.
? Imperceptible: x0 sounds so similar to x that humans cannot differentiate x0 and x when listening to them.
? Robust: x0 is still effective when played by a speaker and recorded by a microphone in an over-the-air attack.
ASR MODEL
最先進(jìn)的 Lingvo classi?er 。
THREAT MODEL
the white box threat model (白盒攻擊模型)
創(chuàng)新點(diǎn):不需要知道攻擊目標(biāo)的準(zhǔn)確配置,而是了解其分布,以便對(duì)抗樣本對(duì)此類分布的攻擊目標(biāo)都有效。
總結(jié)
以上是生活随笔為你收集整理的“Imperceptible,Robust,and Targeted Adversaria lExamples for Automatic Speech Recognition”的全部?jī)?nèi)容,希望文章能夠幫你解決所遇到的問題。
- 上一篇: 【精彩实况】第三届大数据科学与工程国际会
- 下一篇: 热烈祝贺《大数据》主编郑纬民教授当选中国