PrivacyMic: Utilizing Inaudible Frequencies for Privacy Preserving Daily Activity Recognition

Title of the Paper

PrivacyMic: Utilizing Inaudible Frequencies for Privacy Preserving Daily Activity Recognition

Paper Information

• Domain: Privacy protection, ultrasonic frequency, daily activity recognition
• Keywords: acoustics, ultrasound, IoT, privacy protection, smart environments, audio sensing, ubiquitous computing, privacy awareness

Research Background and Problem

• Identified Problems or Challenges:

  • Existing microphones are designed based on the human auditory range (20Hz to 20kHz), which leads to the capture of sensitive voice data, potentially violating user privacy. Additionally, these microphones overlook significant acoustic data beyond the range of human hearing.
  • The current "all-or-nothing" microphone usage model sacrifices useful sound recognition functionality in the name of privacy protection.

• Significance:

  • With the proliferation of voice assistants and smart devices, concerns about audio privacy violations are increasing. Addressing this issue can enhance user trust in smart audio devices while retaining their functionality.

• Research Motivation and Related Work:

  • Previous research has primarily focused on speech analysis or activity recognition using traditional audible frequency data, with little attention paid to infrasound or ultrasound frequencies.
  • Some studies have explored privacy protection, but these methods are mostly software-based filters that are vulnerable to attacks or circumvention.

Solution

• Proposed Solution: PrivacyMic is a Raspberry Pi-based device capable of capturing and utilizing infrasound and ultrasound frequencies for privacy-preserving daily activity recognition. The device filters out voice or all audible frequencies through hardware implementation.

• Innovations:

  • A hardware-level filtering mechanism is designed to remove sensitive voice frequencies before the audio signal passes through the analog-to-digital converter (ADC), fundamentally preventing voice privacy breaches.
  • Acoustic data beyond the human auditory range (e.g., ultrasound) is utilized to compensate for the information loss caused by voice removal, enabling efficient privacy-aware activity recognition.

• Implementation Steps and Key Technologies:

  1. Hardware Design:
     • Equipped with a wideband microphone capable of capturing audio in the range of 0.05Hz to 192kHz.
     • Designed switchable hardware low-pass and high-pass filters to remove voice (below 8kHz) or audible sound (below 16kHz).
  2. Data Collection and Feature Analysis:
     • Collected sound wave data from 127 common household and office devices, generating high-resolution Fast Fourier Transform (FFT) features.
     • Applied information power metrics to evaluate the predictive capability of different frequency bands.
  3. Machine Learning Model Training:
     • Trained a Random Forest classifier to evaluate activity recognition accuracy under different frequency combinations.
  4. Privacy and Performance Validation:
     • Conducted user tests to verify the unintelligibility of filtered voice.
     • Tested the device's recognition accuracy in real-world environments such as kitchens, bathrooms, and offices.

Research Outcomes

• Specific Results:

  • Spectral analysis results show that features in the ultrasonic range (above 16kHz) have the highest average importance for object recognition. Five features in this range rank among the top 10 in information power.
  • PrivacyMic achieves a classification accuracy of 91.4% under privacy-preserving settings with voice filtering and multi-band combinations, significantly outperforming the 50.5% accuracy achieved with voice filtering alone.
  • In real-world tests conducted in kitchens, bathrooms, and office environments, classification accuracy exceeded 95% regardless of the filter settings used.

• Advantages Over Existing Solutions:

  • Hardware-level voice filtering ensures that the device does not transmit sensitive voice information in any scenario, enhancing system security and privacy protection.
  • High activity recognition accuracy is achieved without relying on voice data, expanding the application scope of acoustic signals.

• Experiments and Evaluation Results:

  • In privacy evaluation experiments, participants were unable to understand or transcribe any voice content from the filtered audio.
  • Natural language processing tools could not extract voice information from the filtered audio, confirming the device's privacy protection effectiveness.
  • Real-world activity recognition experiments demonstrated PrivacyMic's stability and high accuracy across different scenarios.

• Limitations and Future Directions:

  • The device does not fully prevent potential future algorithmic voice reconstruction attacks, requiring further improvement in filter cutoff frequencies to reduce risks.
  • Future work could focus on optimizing hardware costs and extending functionality to more scenarios, such as wearable devices or more complex environments.