Gyrophone: Recognizing Speech From Gyroscope Signals

Presentation on theme: "Gyrophone: Recognizing Speech From Gyroscope Signals"— Presentation transcript:

1 Gyrophone: Recognizing Speech From Gyroscope Signals
此页可以删除 狄念 2019年4月

2 MEMS GYROSCOPES

3 MEMS GYROSCOPES (a) STMMicroelectronics (b) InvenSense
MEMS陀螺仪的工作原理是利用科里奥利力，即旋转物体在有径向运动时所收到的切向力。如果物体在圆盘上没有径向运动，科里奥利力就不会产生。因此，在MEMS陀螺仪的设计上，这个物体被驱动，不停地来回做径向运动或者震荡，与此对应的科里奥利力就是不停地在横向来回变化，并有可能使物体在横向作微小震荡，相位正好与驱动力差90度。（图二）MEMS陀螺仪通常有两个方向的可移动电容板。径向的电容板加震荡电压迫使物体作径向运动（有点像加速度计中的自测试模式），横向的电容板测量由于横向科里奥利运动带来的电容变化（就像加速度计测量加速度）。因为科里奥利力正比于角速度，所以由电容的变化可以计算出角速度。 (a) STMMicroelectronics (b) InvenSense

4 SUSCEPTIBLE TO SOUND Audible signal frequency: >20Hz
70 HZ TONE POWER SPECTRAL DENSITY 50 HZ TONE POWER SPECTRAL DENSITY Audible signal frequency: >20Hz mobile device's angular velocity: <20 cycles/s

5 16bit in the latest generations of gyroscopes
Characteristics Sampling resolution: 16bit in the latest generations of gyroscopes Hardware sampling frequency: STM Microelectronic:800Hz InvenSense: up to 8000Hz POTS: 8000Hz Software sampling frequency: Android:200Hz IOS:100Hz Aliasing Self Noise: ~75dB Directionality

6 Experiment-Single Gyroscope
Target: speaker identification(gender) and isolated words recognition Speaker independent/dependent recognition Setup: A sub-woofer, two tweeters(75dB, playback) Subset of TIDigits speech: 10 speakers * 11 words * 2 = 220 recordings

7 Sphinx WAV/Silence removal Tested for isolated word recognition
Lower than 100Hz Rate Human speech ~40% Using a gyroscope 14% Random 9% Tested for isolated word recognition

8 Custom recognition algorithms
SVM GMM DTW Nexus 4 80% 72% 84% Galaxy S Ⅲ 82% 68% 58% Speaker's gender identification results SVM GMM DTW Nexus4 Mixed female/male 23% 21% 50% Female speakers 33% 32% 45% Male speakers 38% 26% 65% Galaxy S Ⅲ 20% 19% 17% 30% 29% 25% Speaker identification results(five speakers)

9 Custom recognition algorithms
SVM GMM DTW Nexus4 Mixed female/male 23% 21% 50% Female speakers 33% 32% 45% Male speakers 38% 26% 65% Galaxy S Ⅲ 20% 19% 17% 30% 29% 25% Speaker-independent case-isolated words recognition results SVM GMM DTW Nexus4 15% 5% 65% Speaker-dependent case-isolated words recognition (single speaker/44 recorded words)

10 Experiment-Multiple Devices

11 Experiment-Multiple Devices
SVM GMM DTW Nexus4 18%(15%) 14%(5%) 77%(65%) Speaker-dependent case-isolated words recognition (two Nexus4 devices/44 recorded words)

12 Further Attacks/Defenses
Increasing the gyro's sampling rate. Source separation Ambient sound recognition Defenses: Low-pass filter the raw samples (enough to pass 0-20Hz frequencies) Access to high sampling rate should require a special permission Hardware level filters for root access Acoustic masking

13 Conclusion The unmitigated access to the gyroscopes may
lead to a leak of privacy, which is dangerous.

14 谢 谢！