1570478458.rar_VoiceAuthentication_语音欺骗

<html xmlns="http://www.w3.org/1999/xhtml"> <head> <meta charset="utf-8"> <meta name="generator" content="pdf2htmlEX"> <meta http-equiv="X-UA-Compatible" content="IE=edge,chrome=1"> <link rel="stylesheet" href="https://static.pudn.com/base/css/base.min.css"> <link rel="stylesheet" href="https://static.pudn.com/base/css/fancy.min.css"> <link rel="stylesheet" href="https://static.pudn.com/prod/directory_preview_static/625b94e3be9ad24cfa1d841d/raw.css"> <script src="https://static.pudn.com/base/js/compatibility.min.js"></script> <script src="https://static.pudn.com/base/js/pdf2htmlEX.min.js"></script> <script> try{ pdf2htmlEX.defaultViewer = new pdf2htmlEX.Viewer({}); }catch(e){} </script> <title></title> </head> <body> <div id="sidebar" style="display: none"> <div id="outline"> </div> </div> <div id="pf1" class="pf w0 h0" data-page-no="1"><div class="pc pc1 w0 h0"><img class="bi x0 y0 w1 h1" alt="" src="https://static.pudn.com/prod/directory_preview_static/625b94e3be9ad24cfa1d841d/bg1.jpg"><div class="t m0 x1 h2 y1 ff1 fs0 fc0 sc0 ls0 ws0">VoicePop: A Pop Noise based Anti-spooﬁng System</div><div class="t m0 x2 h2 y2 ff1 fs0 fc0 sc0 ls0 ws0">for Voice Authentication on Smartphones</div><div class="t m0 x3 h3 y3 ff1 fs1 fc0 sc0 ls0 ws0">Qian Wang</div><div class="t m0 x4 h4 y4 ff2 fs2 fc0 sc0 ls0 ws0">†</div><div class="t m0 x5 h3 y3 ff1 fs1 fc0 sc0 ls0 ws0">, Xiu Lin</div><div class="t m0 x6 h4 y4 ff2 fs2 fc0 sc0 ls0 ws0">†</div><div class="t m0 x7 h3 y3 ff1 fs1 fc0 sc0 ls0 ws0">, Man Zhou</div><div class="t m0 x8 h4 y4 ff2 fs2 fc0 sc0 ls0 ws0">†</div><div class="t m0 x9 h3 y3 ff1 fs1 fc0 sc0 ls0 ws0">, Yanjiao Chen</div><div class="t m0 xa h4 y4 ff2 fs2 fc0 sc0 ls0 ws0">∗</div><div class="t m0 xb h3 y3 ff1 fs1 fc0 sc0 ls0 ws0">, Cong Wang</div><div class="t m0 xc h4 y4 ff2 fs2 fc0 sc0 ls0 ws0">‡</div><div class="t m0 xd h3 y3 ff1 fs1 fc0 sc0 ls0 ws0">, Qi Li</div><div class="t m0 xe h4 y4 ff2 fs2 fc0 sc0 ls0 ws0">§</div><div class="t m0 xf h3 y3 ff1 fs1 fc0 sc0 ls0 ws0">, Xiangyang Luo</div><div class="t m0 x10 h4 y4 ff2 fs2 fc0 sc0 ls0 ws0">¶</div><div class="t m0 x4 h4 y5 ff2 fs2 fc0 sc0 ls0 ws0">†</div><div class="t m0 x5 h5 y6 ff1 fs3 fc0 sc0 ls0 ws0">School of Cyber Science and Engineering, Wuhan University, P. R. China.</div><div class="t m0 x11 h4 y7 ff2 fs2 fc0 sc0 ls0 ws0">∗</div><div class="t m0 x12 h5 y8 ff1 fs3 fc0 sc0 ls0 ws0">School of Computer Science, Wuhan University, P. R. China.</div><div class="t m0 x13 h4 y9 ff2 fs2 fc0 sc0 ls0 ws0">‡</div><div class="t m0 x14 h5 ya ff1 fs3 fc0 sc0 ls0 ws0">Department of Computer Science, City University of Hong Kong, Hong Kong, P. R. China.</div><div class="t m0 x15 h4 yb ff2 fs2 fc0 sc0 ls0 ws0">§</div><div class="t m0 x16 h5 yc ff1 fs3 fc0 sc0 ls0 ws0">Institute for Network Sciences and Cyberspace, Tsinghua University, P. R. China.</div><div class="t m0 x17 h4 yd ff2 fs2 fc0 sc0 ls0 ws0">¶</div><div class="t m0 x18 h5 ye ff1 fs3 fc0 sc0 ls0 ws0">The State Key Laboratory of Mathematical Engineering and Advanced Computing, Zhengzhou, P. R. China.</div><div class="t m0 x19 h6 yf ff3 fs4 fc0 sc0 ls0 ws0">Abstract—Voice biometrics is widely adopted for identity</div><div class="t m0 x1 h7 y10 ff4 fs4 fc0 sc0 ls0 ws0">authentication in mobile devices. However, voice authentication</div><div class="t m0 x1 h7 y11 ff4 fs4 fc0 sc0 ls0 ws0">is vulnerable to spooﬁng attacks, where an adversary may</div><div class="t m0 x1 h7 y12 ff4 fs4 fc0 sc0 ls0 ws0">deceive the voice authentication system with pre-recorded or</div><div class="t m0 x1 h7 y13 ff4 fs4 fc0 sc0 ls0 ws0">synthesized samples from the legitimate user or by impersonating</div><div class="t m0 x1 h7 y14 ff4 fs4 fc0 sc0 ls0 ws0">the speaking style of the targeted user. In this paper, we design</div><div class="t m0 x1 h7 y15 ff4 fs4 fc0 sc0 ls0 ws0">and implement VoicePop, a robust software-only anti-spooﬁng</div><div class="t m0 x1 h6 y16 ff4 fs4 fc0 sc0 ls0 ws0">system on smartphones. VoicePop leverages the pop noise, which</div><div class="t m0 x1 h7 y17 ff4 fs4 fc0 sc0 ls0 ws0">is produced by the user breathing while speaking close to the</div><div class="t m0 x1 h7 y18 ff4 fs4 fc0 sc0 ls0 ws0">microphone. The pop noise is delicate and subject to user</div><div class="t m0 x1 h7 y19 ff4 fs4 fc0 sc0 ls0 ws0">diversity, making it hard to record by replay attacks beyond</div><div class="t m0 x1 h7 y1a ff4 fs4 fc0 sc0 ls0 ws0">a certain distance and to imitate precisely by impersonators.</div><div class="t m0 x1 h7 y1b ff4 fs4 fc0 sc0 ls0 ws0">We design a novel pop noise detection scheme to pinpoint pop</div><div class="t m0 x1 h7 y1c ff4 fs4 fc0 sc0 ls0 ws0">noises at the phonemic level, based on which we establish indi-</div><div class="t m0 x1 h7 y1d ff4 fs4 fc0 sc0 ls0 ws0">vidually unique relationship between phonemes and pop noises</div><div class="t m0 x1 h7 y1e ff4 fs4 fc0 sc0 ls0 ws0">to identify legitimate users and defend against spooﬁng attacks.</div><div class="t m0 x1 h7 y1f ff4 fs4 fc0 sc0 ls0 ws0">Our experimental results with 18 participants and three types of</div><div class="t m0 x1 h7 y20 ff4 fs4 fc0 sc0 ls0 ws0">smartphones show that VoicePop achieves over 93.5% detection</div><div class="t m0 x1 h7 y21 ff4 fs4 fc0 sc0 ls0 ws0">accuracy at around 5.4% equal error rate. VoicePop requires</div><div class="t m0 x1 h7 y22 ff4 fs4 fc0 sc0 ls0 ws0">no additional hardware but only the built-in microphones in</div><div class="t m0 x1 h7 y23 ff4 fs4 fc0 sc0 ls0 ws0">virtually all smartphones, which can be readily integrated in</div><div class="t m0 x1 h7 y24 ff4 fs4 fc0 sc0 ls0 ws0">existing voice authentication systems for mobile devices.</div><div class="t m0 x1a h5 y25 ff1 fs3 fc0 sc0 ls0 ws0">I. INTRODUCTION</div><div class="t m0 x19 h5 y26 ff1 fs3 fc0 sc0 ls0 ws0">Compared with password-based authentication, voice au-</div><div class="t m0 x1 h5 y27 ff1 fs3 fc0 sc0 ls0 ws0">thentication is more convenient since it is hands-free and users</div><div class="t m0 x1 h5 y28 ff1 fs3 fc0 sc0 ls0 ws0">do not need to memorize passwords. In recent years, the rapid</div><div class="t m0 x1 h5 y29 ff1 fs3 fc0 sc0 ls0 ws0">growth of mobile communications has boosted the use of voice</div><div class="t m0 x1 h5 y2a ff1 fs3 fc0 sc0 ls0 ws0">authentication in mobile devices, including smartphone login,</div><div class="t m0 x1 h5 y2b ff1 fs3 fc0 sc0 ls0 ws0">mobile banking and e-commerce. For example, Google allows</div><div class="t m0 x1 h5 y2c ff1 fs3 fc0 sc0 ls0 ws0">users to unlock their phones of Android operating systems by</div><div class="t m0 x1 h5 y2d ff1 fs3 fc0 sc0 ls0 ws0">voice biometrics [1]. Say Tec uses voice biometric solution to</div><div class="t m0 x1 h5 y2e ff1 fs3 fc0 sc0 ls0 ws0">support mobile ﬁnancial services such as online payment and</div><div class="t m0 x1 h5 y2f ff1 fs3 fc0 sc0 ls0 ws0">banking [2].</div><div class="t m0 x19 h5 y30 ff1 fs3 fc0 sc0 ls0 ws0">However, since the sound transmits through an open and</div><div class="t m0 x1 h5 y31 ff1 fs3 fc0 sc0 ls0 ws0">public channel, the voice authentication system is highly vul-</div><div class="t m0 x1 h5 y32 ff1 fs3 fc0 sc0 ls0 ws0">nerable to spooﬁng attacks [3]–[5]. There are two major types</div><div class="t m0 x1 h5 y33 ff1 fs3 fc0 sc0 ls0 ws0">of spooﬁng attacks, namely replay attacks and impersonation</div><div class="t m0 x1 h5 y34 ff1 fs3 fc0 sc0 ls0 ws0">attacks [6]. In replay attacks, the adversary pre-records and</div><div class="t m0 x1 h5 y35 ff1 fs3 fc0 sc0 ls0 ws0">playbacks the voice sample of the passphrase of a legal user</div><div class="t m0 x1 h5 y36 ff1 fs3 fc0 sc0 ls0 ws0">to deceive the authentication system [7]. An adversary can</div><div class="t m0 x1 h5 y37 ff1 fs3 fc0 sc0 ls0 ws0">also mimic the voice characteristics and style of a legal user</div><div class="t m0 x1 h5 y38 ff1 fs3 fc0 sc0 ls0 ws0">to conduct impersonation attacks [8]. Spooﬁng attacks may</div><div class="t m0 x1 h5 y39 ff1 fs3 fc0 sc0 ls0 ws0">greatly harm the users as the adversary may gain access to the</div><div class="t m0 x1b h5 yf ff1 fs3 fc0 sc0 ls0 ws0">victim’s smartphone to steal private information and perform</div><div class="t m0 x1b h5 y3a ff1 fs3 fc0 sc0 ls0 ws0">malicious operations.</div><div class="t m0 x1c h5 y3b ff1 fs3 fc0 sc0 ls0 ws0">Traditional methods to defend against replay attacks and</div><div class="t m0 x1b h5 y3c ff1 fs3 fc0 sc0 ls0 ws0">impersonation attacks are liveness detection and automatic</div><div class="t m0 x1b h5 y3d ff1 fs3 fc0 sc0 ls0 ws0">speaker veriﬁcation (ASV) system. Liveness detection exam-</div><div class="t m0 x1b h5 y15 ff1 fs3 fc0 sc0 ls0 ws0">ines whether the voice is produced by a live user or a speaker,</div><div class="t m0 x1b h5 y3e ff1 fs3 fc0 sc0 ls0 ws0">and ASV leverages unique spectral and prosodic features of the</div><div class="t m0 x1b h5 y3f ff1 fs3 fc0 sc0 ls0 ws0">user’s voice for identity authentication. For example, Zhang et</div><div class="t m0 x1b h5 y40 ff5 fs3 fc0 sc0 ls0 ws0">al. [9] proposed to capture time-difference-of-arrival (TDoA)</div><div class="t m0 x1b h5 y41 ff1 fs3 fc0 sc0 ls0 ws0">changes to the two microphones of the phone in a sequence</div><div class="t m0 x1b h5 y1b ff1 fs3 fc0 sc0 ls0 ws0">of phoneme sounds to differentiate the voice from a live user</div><div class="t m0 x1b h5 y42 ff1 fs3 fc0 sc0 ls0 ws0">and a replay device, but the user has to hold the phone at a</div><div class="t m0 x1b h5 y43 ff1 fs3 fc0 sc0 ls0 ws0">speciﬁc position. In [10], the smartphone served as a Doppler</div><div class="t m0 x1b h5 y44 ff1 fs3 fc0 sc0 ls0 ws0">radar to transmit a high-frequency acoustic sound from the</div><div class="t m0 x1b h5 y45 ff1 fs3 fc0 sc0 ls0 ws0">built-in speaker and monitor the reﬂections of articulators</div><div class="t m0 x1b h5 y46 ff1 fs3 fc0 sc0 ls0 ws0">at the microphone for liveness detection. Unfortunately, the</div><div class="t m0 x1b h5 y47 ff1 fs3 fc0 sc0 ls0 ws0">extent of articulatory movements affects the effectiveness of</div><div class="t m0 x1b h5 y48 ff1 fs3 fc0 sc0 ls0 ws0">this countermeasure. Chen et al. [11] explored the magnetic</div><div class="t m0 x1b h5 y49 ff1 fs3 fc0 sc0 ls0 ws0">ﬁeld emitted from loudspeakers to detect voice replay attacks.</div><div class="t m0 x1b h5 y4a ff1 fs3 fc0 sc0 ls0 ws0">However, users need to move the smartphone with a predeﬁned</div><div class="t m0 x1b h5 y4b ff1 fs3 fc0 sc0 ls0 ws0">trajectory around the mouth while speaking the passphrase.</div><div class="t m0 x1b h5 y4c ff1 fs3 fc0 sc0 ls0 ws0">M Sahidullah et al. [12] developed an ASV system against</div><div class="t m0 x1b h5 y4d ff1 fs3 fc0 sc0 ls0 ws0">impersonation attacks using the throat microphone which is</div><div class="t m0 x1b h5 y4e ff1 fs3 fc0 sc0 ls0 ws0">not available in most smartphones.</div><div class="t m0 x1c h5 y4f ff1 fs3 fc0 sc0 ls0 ws0">In this paper, we propose and implement VoicePop, a novel</div><div class="t m0 x1b h5 y50 ff1 fs3 fc0 sc0 ls0 ws0">and practical anti-spooﬁng system based on pop noise that is</div><div class="t m0 x1b h5 y51 ff1 fs3 fc0 sc0 ls0 ws0">induced by the user breathing while speaking the passphrase</div><div class="t m0 x1b h5 y52 ff1 fs3 fc0 sc0 ls0 ws0">close to the microphone. The recorded voice samples hardly</div><div class="t m0 x1b h5 y53 ff1 fs3 fc0 sc0 ls0 ws0">contain the pop noise since the sound of breath is gentle</div><div class="t m0 x1b h5 y54 ff1 fs3 fc0 sc0 ls0 ws0">compared to the speech and will die out beyond a certain</div><div class="t m0 x1b h5 y55 ff1 fs3 fc0 sc0 ls0 ws0">distance. The pop noise is also subject to user diversity and it</div><div class="t m0 x1b h5 y30 ff1 fs3 fc0 sc0 ls0 ws0">is very difﬁcult for attackers to imitate the way of breathing</div><div class="t m0 x1b h5 y56 ff1 fs3 fc0 sc0 ls0 ws0">of the legal user. These ideal properties of the pop noise</div><div class="t m0 x1b h5 y32 ff1 fs3 fc0 sc0 ls0 ws0">enable our proposed VoicePop system to resist spooﬁng attacks</div><div class="t m0 x1b h5 y33 ff1 fs3 fc0 sc0 ls0 ws0">in voice authentication. To begin with, we conduct phoneme</div><div class="t m0 x1b h5 y34 ff1 fs3 fc0 sc0 ls0 ws0">segmentation on the collected voice sample according to spec-</div><div class="t m0 x1b h5 y35 ff1 fs3 fc0 sc0 ls0 ws0">trogram characteristics. We design a novel pop noise detection</div><div class="t m0 x1b h5 y36 ff1 fs3 fc0 sc0 ls0 ws0">algorithm to locate pop noises at the phonemic level. A lack of</div><div class="t m0 x1b h5 y57 ff1 fs3 fc0 sc0 ls0 ws0">pop noise is deemed as a replay attack, however, environmental</div><div class="t m0 x1b h5 y38 ff1 fs3 fc0 sc0 ls0 ws0">noise and hardware noise may also be wrongly detected as pop</div><div class="t m0 x1b h5 y39 ff1 fs3 fc0 sc0 ls0 ws0">noises in the replayed voice samples. To address this problem,</div></div><div class="pi" data-data='{"ctm":[1.568627,0.000000,0.000000,1.568627,0.000000,0.000000]}'></div></div> </body> </html>

<div id="pf2" class="pf w0 h0" data-page-no="2"><div class="pc pc2 w0 h0"><img class="bi x0 y0 w1 h1" alt="" src="https://static.pudn.com/prod/directory_preview_static/625b94e3be9ad24cfa1d841d/bg2.jpg"><div class="t m0 x1 h5 y58 ff1 fs3 fc0 sc0 ls0 ws0">we extract the Gammatone Frequency Cepstral Coefﬁcients</div><div class="t m0 x1 h5 y59 ff1 fs3 fc0 sc0 ls0 ws0">(GFCC) features of the detected pop noises for classiﬁcation</div><div class="t m0 x1 h5 y5a ff1 fs3 fc0 sc0 ls0 ws0">to distinguish a genuine voice sample and a replayed one.</div><div class="t m0 x1 h5 y5b ff1 fs3 fc0 sc0 ls0 ws0">To defend against impersonation attacks, we leverage the</div><div class="t m0 x1 h5 y5c ff1 fs3 fc0 sc0 ls0 ws0">individually unique relationship between phonemes and pop</div><div class="t m0 x1 h5 y5d ff1 fs3 fc0 sc0 ls0 ws0">noises to construct a phoneme-pop sequence. A legal user is</div><div class="t m0 x1 h5 y5e ff1 fs3 fc0 sc0 ls0 ws0">accepted if the phoneme-pop sequence of the voice sample is</div><div class="t m0 x1 h5 y5f ff1 fs3 fc0 sc0 ls0 ws0">similar to that stored in the user proﬁle upon registration, and</div><div class="t m0 x1 h5 y60 ff1 fs3 fc0 sc0 ls0 ws0">an impersonation attack is declared otherwise.</div><div class="t m0 x19 h5 y61 ff1 fs3 fc0 sc0 ls0 ws0">VoicePop requires no additional hardware but only the</div><div class="t m0 x1 h5 y62 ff1 fs3 fc0 sc0 ls0 ws0">built-in microphones that are available on almost all mobile</div><div class="t m0 x1 h5 y63 ff1 fs3 fc0 sc0 ls0 ws0">devices. VoicePop also demands no extra efforts from users</div><div class="t m0 x1 h5 y64 ff1 fs3 fc0 sc0 ls0 ws0">except to speak the passphrase as required by current voice</div><div class="t m0 x1 h5 y65 ff1 fs3 fc0 sc0 ls0 ws0">authentication systems. As far as we are concerned, we are</div><div class="t m0 x1 h5 y66 ff1 fs3 fc0 sc0 ls0 ws0">the ﬁrst to use the features of pop noise to defend both replay</div><div class="t m0 x1 h5 y67 ff1 fs3 fc0 sc0 ls0 ws0">attacks and impersonation attacks. We implement VoicePop on</div><div class="t m0 x1 h5 y68 ff1 fs3 fc0 sc0 ls0 ws0">3 types of smartphones and evaluate its performance with 18</div><div class="t m0 x1 h5 y69 ff1 fs3 fc0 sc0 ls0 ws0">volunteers under different experimental settings. The results</div><div class="t m0 x1 h5 y6a ff1 fs3 fc0 sc0 ls0 ws0">verify the effectiveness of VoicePop that achieves over 93.5%</div><div class="t m0 x1 h5 y6b ff1 fs3 fc0 sc0 ls0 ws0">detection accuracy at around 5.4% equal error rate. The main</div><div class="t m0 x1 h5 y6c ff1 fs3 fc0 sc0 ls0 ws0">contributions of this work are summarized as follows:</div><div class="t m0 x19 h5 y6d ff6 fs5 fc0 sc0 ls0 ws0">• We propose VoicePop, a practical and effective software-</div><div class="t m0 x1d h5 y6e ff1 fs3 fc0 sc0 ls0 ws0">only anti-spooﬁng system for voice authentication based</div><div class="t m0 x1d h5 y6f ff1 fs3 fc0 sc0 ls0 ws0">on pop noise, which can be easily integrated in commer-</div><div class="t m0 x1d h5 y70 ff1 fs3 fc0 sc0 ls0 ws0">cial off-the-shelf smartphones.</div><div class="t m0 x19 h5 y71 ff6 fs5 fc0 sc0 ls0 ws0">• We design a novel pop noise detection scheme to de-</div><div class="t m0 x1d h5 y72 ff1 fs3 fc0 sc0 ls0 ws0">fend against replay spooﬁng attacks, and leverage the</div><div class="t m0 x1d h5 y73 ff1 fs3 fc0 sc0 ls0 ws0">individually unique relationship between phonemes and</div><div class="t m0 x1d h5 y74 ff1 fs3 fc0 sc0 ls0 ws0">pop noises to generate a phoneme-pop sequence to resist</div><div class="t m0 x1d h5 y75 ff1 fs3 fc0 sc0 ls0 ws0">impersonation spooﬁng attacks.</div><div class="t m0 x19 h5 y76 ff6 fs5 fc0 sc0 ls0 ws0">• We build a fully-functional VoicePop prototype using</div><div class="t m0 x1d h5 y77 ff1 fs3 fc0 sc0 ls0 ws0">off-the-shelf smartphones. Extensive evaluation results</div><div class="t m0 x1d h5 y78 ff1 fs3 fc0 sc0 ls0 ws0">demonstrate that VoicePop can detect both replay and</div><div class="t m0 x1d h5 y79 ff1 fs3 fc0 sc0 ls0 ws0">impersonation spooﬁng attacks with a high accuracy and</div><div class="t m0 x1d h5 y7a ff1 fs3 fc0 sc0 ls0 ws0">a low equal error rate.</div><div class="t m0 x1e h5 y7b ff1 fs3 fc0 sc0 ls0 ws0">II. PRELIMINARIES</div><div class="t m0 x1 h8 y7c ff5 fs3 fc0 sc0 ls0 ws0">A. Attack Model</div><div class="t m0 x19 h5 y7d ff1 fs3 fc0 sc0 ls0 ws0">Voice authentication system can be text-dependent (requires</div><div class="t m0 x1 h5 y7e ff1 fs3 fc0 sc0 ls0 ws0">the same password for enrolment and veriﬁcation) or text-</div><div class="t m0 x1 h5 y7f ff1 fs3 fc0 sc0 ls0 ws0">independent (accept arbitrary utterances from speakers). We</div><div class="t m0 x1 h5 y80 ff1 fs3 fc0 sc0 ls0 ws0">primarily focus on the text-dependent authentication system,</div><div class="t m0 x1 h5 y81 ff1 fs3 fc0 sc0 ls0 ws0">which is currently the most widely adopted and commercially</div><div class="t m0 x1 h5 y82 ff1 fs3 fc0 sc0 ls0 ws0">viable method with a high authentication accuracy [13]. Fig. 1</div><div class="t m0 x1 h5 y83 ff1 fs3 fc0 sc0 ls0 ws0">displays a typical voice authentication system. For the attack</div><div class="t m0 x1 h5 y84 ff1 fs3 fc0 sc0 ls0 ws0">model, we consider replay spooﬁng attacks and impersonation</div><div class="t m0 x1 h5 y85 ff1 fs3 fc0 sc0 ls0 ws0">spooﬁng attacks.</div><div class="t m0 x19 h9 y34 ff4 fs3 fc0 sc0 ls0 ws0">Replay attacks. Replay attacks leverage computers and</div><div class="t m0 x1 h5 y35 ff1 fs3 fc0 sc0 ls0 ws0">other peripheral devices (e.g., loudspeaker) to perform voice</div><div class="t m0 x1 h5 y36 ff1 fs3 fc0 sc0 ls0 ws0">playback to the microphone of the smartphone. The replay</div><div class="t m0 x1 h5 y37 ff1 fs3 fc0 sc0 ls0 ws0">samples that involve the information of the victim’s passphrase</div><div class="t m0 x1 h5 y38 ff1 fs3 fc0 sc0 ls0 ws0">can be produced by stealthily recording, voice synthesis,</div><div class="t m0 x1 h5 y39 ff1 fs3 fc0 sc0 ls0 ws0">and voice conversion. In this paper, we mainly focus on</div><div class="c x1f y86 w2 ha"><div class="t m0 x20 hb y87 ff7 fs6 fc0 sc0 ls0 ws0">Microphone</div><div class="t m0 x21 hc y88 ff8 fs6 fc0 sc0 ls0 ws0">Voice </div><div class="t m0 x22 hc y89 ff8 fs6 fc0 sc0 ls0 ws0">Data</div><div class="t m0 x23 hc y8a ff8 fs6 fc0 sc0 ls0 ws0">User</div><div class="t m0 x24 hb y8b ff7 fs6 fc0 sc0 ls0 ws0">Feature </div><div class="t m0 x25 hb y8c ff7 fs6 fc0 sc0 ls0 ws0">Extraction</div><div class="t m0 x26 hb y87 ff7 fs6 fc0 sc0 ls0 ws0">Classifier</div><div class="t m0 x27 hc y8b ff8 fs6 fc0 sc0 ls0 ws0">Voice </div><div class="t m0 x28 hc y8c ff8 fs6 fc0 sc0 ls0 ws0">Feature</div><div class="t m0 x29 hc y8d ff8 fs6 fc0 sc0 ls0 ws0">Speaker Model</div><div class="t m0 x2a hc y8e ff8 fs6 fc0 sc0 ls0 ws0">Decision</div><div class="t m0 x2b hc y8f ff8 fs6 fc0 sc0 ls0 ws0">Reference</div></div><div class="t m0 x2c hd y90 ff1 fs2 fc0 sc0 ls0 ws0">Fig. 1. A typical voice authentication system.</div><div class="t m0 x1b h5 y91 ff1 fs3 fc0 sc0 ls0 ws0">replay attacks by pre-recording since they retain more user</div><div class="t m0 x1b h5 y92 ff1 fs3 fc0 sc0 ls0 ws0">characteristics than those generated by synthesis or conversion.</div><div class="t m0 x1c h9 y93 ff4 fs3 fc0 sc0 ls0 ws0">Impersonation attacks. Impersonation attacks can be con-</div><div class="t m0 x1b h5 y94 ff1 fs3 fc0 sc0 ls0 ws0">ducted in two ways. The ﬁrst is simply to imitate the legitimate</div><div class="t m0 x1b h5 y95 ff1 fs3 fc0 sc0 ls0 ws0">user’s voice and speaking habit without the help of other de-</div><div class="t m0 x1b h5 y96 ff1 fs3 fc0 sc0 ls0 ws0">vices. The second is more advanced where we consider that the</div><div class="t m0 x1b h5 y97 ff1 fs3 fc0 sc0 ls0 ws0">attacker knows the key rationale of our anti-spooﬁng system</div><div class="t m0 x1b h5 y98 ff1 fs3 fc0 sc0 ls0 ws0">and observes how the target user pronounces the passphrase.</div><div class="t m0 x1b h5 y99 ff1 fs3 fc0 sc0 ls0 ws0">To perform this type of attacks, we assume that the adversary</div><div class="t m0 x1b h5 y9a ff1 fs3 fc0 sc0 ls0 ws0">uses a loudspeaker to replay the pre-recorded voice sample</div><div class="t m0 x1b h5 y9b ff1 fs3 fc0 sc0 ls0 ws0">near the microphone while simultaneously impersonating the</div><div class="t m0 x1b h5 y9c ff1 fs3 fc0 sc0 ls0 ws0">victim’s breathing pattern closely to the microphone.</div><div class="t m0 x1b h8 y9d ff5 fs3 fc0 sc0 ls0 ws0">B. Pop Noise</div><div class="t m0 x1c h5 y9e ff1 fs3 fc0 sc0 ls0 ws0">The human voice is produced through several stages. Air</div><div class="t m0 x1b h5 y9f ff1 fs3 fc0 sc0 ls0 ws0">is ﬁrst expelled from the lung to form an airﬂow, which then</div><div class="t m0 x1b h5 ya0 ff1 fs3 fc0 sc0 ls0 ws0">enters the throat, passes through the vocal cords into the vocal</div><div class="t m0 x1b h5 ya1 ff1 fs3 fc0 sc0 ls0 ws0">tract, and ﬁnally bursts out of the mouth to form the sound</div><div class="t m0 x1b h5 ya2 ff1 fs3 fc0 sc0 ls0 ws0">wave. When the resulting airﬂow reaches the microphone,</div><div class="t m0 x1b h5 ya3 ff1 fs3 fc0 sc0 ls0 ws0">if the user’s mouth is close enough to the microphone, the</div><div class="t m0 x1b h5 ya4 ff1 fs3 fc0 sc0 ls0 ws0">captured sound signals will not only contain the speech</div><div class="t m0 x1b h5 ya5 ff1 fs3 fc0 sc0 ls0 ws0">information but also the plosive burst as the friction between</div><div class="t m0 x1b h5 ya6 ff1 fs3 fc0 sc0 ls0 ws0">the lips and the airﬂow, known as the pop noise. In contrast,</div><div class="t m0 x1b h5 ya7 ff1 fs3 fc0 sc0 ls0 ws0">an attacker who tries to launch a replay attack usually cannot</div><div class="t m0 x1b h5 ya8 ff1 fs3 fc0 sc0 ls0 ws0">put the microphone of the recording device very close to</div><div class="t m0 x1b h5 ya9 ff1 fs3 fc0 sc0 ls0 ws0">the user’s mouth, thus the recorded voice contains no pop</div><div class="t m0 x1b h5 yaa ff1 fs3 fc0 sc0 ls0 ws0">noise. Therefore, by detecting the pop noise, we are able to</div><div class="t m0 x1b h5 yab ff1 fs3 fc0 sc0 ls0 ws0">distinguish the real speech from a live user and the recorded</div><div class="t m0 x1b h5 yac ff1 fs3 fc0 sc0 ls0 ws0">speech from a loudspeaker.</div><div class="t m0 x1c h5 yad ff1 fs3 fc0 sc0 ls0 ws0">To detect pop noise, we compare the spectrograms of speech</div><div class="t m0 x1b h5 yae ff1 fs3 fc0 sc0 ls0 ws0">signals with and without a pop noise ﬁlter using three different</div><div class="t m0 x1b h5 yaf ff1 fs3 fc0 sc0 ls0 ws0">smartphones, as shown in Fig. 2. It can be found that pop noise</div><div class="t m0 x1b he yb0 ff1 fs3 fc0 sc0 ls0 ws0">has a high energy in the low frequency (typically 0∼100 Hz),</div><div class="t m0 x1b h5 yb1 ff1 fs3 fc0 sc0 ls0 ws0">which has been discussed in the prior study [14]. Moreover,</div><div class="t m0 x1b he yb2 ff1 fs3 fc0 sc0 ls0 ws0">the duration of pop noise varies in the range 20∼100 msec</div><div class="t m0 x1b h5 yb3 ff1 fs3 fc0 sc0 ls0 ws0">based on the way people speak and breathe. Our detection</div><div class="t m0 x1b h5 yb4 ff1 fs3 fc0 sc0 ls0 ws0">algorithm is based on these observations.</div><div class="t m0 x1b h8 yb5 ff5 fs3 fc0 sc0 ls0 ws0">C. Phoneme and Pop Noise</div><div class="t m0 x1c h5 y32 ff1 fs3 fc0 sc0 ls0 ws0">A phoneme is the smallest distinctive unit sound of a</div><div class="t m0 x1b h5 y33 ff1 fs3 fc0 sc0 ls0 ws0">language in the human speech production system.There are</div><div class="t m0 x1b h5 y34 ff1 fs3 fc0 sc0 ls0 ws0">two categories of phonemes, the vowel and the consonant.</div><div class="t m0 x1b h5 y35 ff1 fs3 fc0 sc0 ls0 ws0">A vowel is a sound produced by the airﬂow through the</div><div class="t m0 x1b h5 y36 ff1 fs3 fc0 sc0 ls0 ws0">mouth without hindrance, while a consonant is produced by</div><div class="t m0 x1b h5 y57 ff1 fs3 fc0 sc0 ls0 ws0">obstructing the airﬂow out of the mouth with the teeth, tongue,</div><div class="t m0 x1b h5 y38 ff1 fs3 fc0 sc0 ls0 ws0">lips or palate. Since each phoneme features unique physical</div><div class="t m0 x1b h5 y39 ff1 fs3 fc0 sc0 ls0 ws0">origin in the human vocal tract system and has its own manner</div></div><div class="pi" data-data='{"ctm":[1.568627,0.000000,0.000000,1.568627,0.000000,0.000000]}'></div></div>