Skip to content

Latest commit

 

History

History
67 lines (56 loc) · 3.57 KB

seungwook_19.md

File metadata and controls

67 lines (56 loc) · 3.57 KB

SonicPrint: A Generally adoptable and secure fingerprint biometrics in Smart Devices

Aditya Singh Rathore et al., University of Buffalo - MobiSys2020 best paper

Summarized by Seungwook Kim

총평:

  • Coursework에서 Paper critique 해서 읽었습니다 (지능형모바일시스템)
  • 딥러닝과 관련 되어있지만, 새로운 학습적 novelty를 제안한 것은 아닙니다.
  • 일반적인 지문인식을 통한 authentication이 아니라, 손가락과 휴대폰의 상호작용을 통해 발생하는 소리를 통해 사용자를 identify/authenticate하는 논문입니다. 신기한 내용이어서 재미있게 읽었습니다.
  • 음성 전처리/후처리 부분은 잘 몰라서 제대로 읽지 못했습니다.
  • best paper를 받을만한 innovation/실험 설정 / 실험 결과 라고 생각합니다.

Task: Biometric system for user identification

Contributions

  • Hypothesizes that the FiSe(finger-induced sonic effect) from "finger" swiping on "phone" contains intrinsic fingerprint information
  • Design SonicPrint: a biometric system applicable to smart devices using this FiSe
  • Shows up to 97% accuracy on mobile phones, and resilience against fake-finger and replay attacks
  • Note that since this method uses the 'sound' input, it does not require touch sensors - and is (in theory) applicable to all smart devices

Motivation: Adopting touch sensors in small smart devices in infeasible (hardware constraints)

Method outline

  1. Background isolation

    • Pre-process sound input (high pass filter)
    • Sonic effect enhancement using Multi-band spectral subtraction
    • Denoising-aware wavelet reconstruction

  2. Friction event detection

    • Adaptive detection via Hidden Markov Model (acoustic event detection)
    • phase-based detection & duration verificatino for robustness

  3. Acoustic fingerprint analysis

    • Identify fingerprint features (by level) -> then determine fingerprint-induced sonic features
    • ex) Arch, Line-Unit, Pores are fingerprint features
    • ex) Temporal centroid, Mel-frequency cepstral coefficients are fingerprint-induced sonic features

  4. Ensemble classification

    • Uses ensemble of the following models for final identification:
      • Logistic Regression
      • SVM
      • Random Forest
      • Linear Discriminant Analysis
      • Gaussian Mixture Model

Evaluation / Discussions

  • based on data & reponses from 31 users

  • Balanced accuracy around 80% ~ 90% based on the type of interaction (type of swipe)

  • Reaches up to 97% when the surface is rough, and the sound is more distinguishable

  • Successfully eliminates alien fingers (untrained fingers) at test time

    • FP rate around 2%

  • As the number of subjects increase, performance decreases

    • But usually do not register a lot of subjects in one phone
    • After 15 subjects, learns more accurate features for better identification! (performance bounces)

  • more suitable for rough spaces vs smooth spaces

  • The more curved the surface, lower performance

    • since only partial fingerprint is in contact with the surface

  • Performance is proportional to complexity of swipe pattern

    • more complex -> longer length -> more user-specific information for identification

  • Performance also depends on position of swipe w.r.t microphone position

  • Resilient against fingerprint phantom attack (fake finger) and replay attack (record sound and replay)

    • due to different surface characteristics & small sound

  • May be vulnerable to inaudible ultrasound signals

  • Not robust when trained only in controlled environment (with controlled noise)