This document summarizes research on preserving privacy in audio sensing. It presents two approaches: sound shredding and sound subsampling. Sound shredding randomizes audio frames, while subsampling discards some frames. Experiments show these methods slightly reduce context recognition accuracy but significantly impair speech recognition, gender identification, and other privacy risks. However, sound shredding can be partially reconstructed by matching frame frequencies. Overall, the work introduces privacy-preserving audio techniques but recognizes limitations like sound shredding potentially being attacked through reconstruction.

1. Sound Shredding : Privacy
Preserved Audio Sensing
Presenter: Moustafa Alzantot (UCLA)
Sumeet Kumar, et al.
Carnegie Melon University

2. Introduction
 Sound sensing can be very useful for context
awareness.
 Identify user location and activities
 Potential risks on user’s privacy
 Speech recognition
 Speaker identification
 How to preserve user privacy without comprising the
context awareness accuracy ?

3. Research Question
 This paper presents two approaches for
preserving user privacy without significantly
decreasing the context recognition accuracy
or consuming much battery in
Encryption/Decryption.
 Sound shredding
 Sound subsampling

4. Methodology
Activity context: the place where the activity takes place (e.g.
restaurant for dinning)
Context identification process:
 Audio Data Collection:
 35 sounds collected at 8KHz using nexus 4 phone.
 Feature Extraction:
 Sliding window frame (40 ms window , 50%overlap)
 12 MFCC features for every window.
 Context Recognition:
 Experiments using both simple KNN, and SVM.

5. Methodology
 Sound Subsampling: collection part of raw data.
 50% subsampling discarding one frame after every single frame is
stored.
 Subsampling results in a slight drop in context recognition
accuracy.

6. Methodology
 Sound Shredding: randomize the audio
frames order in a sound snippet.

7. Results : Context Recognition
Accuracy
 Collected 35 sound samples in different contexts
(faculty meeting, restaurant, walking, coffee shop)
 80% of data for training, 20% for testing.
 Context recognition accuracy is slightly dropped.

8. Results: Privacy User Study
 User study involves playing different sounds (shredded, and sub-
sampled)
 Users rated the ability of speech recognition, gender identification,
and people counting.
 Scale used from 1(Yes, I can) to 5 (Not, at all).
 Gender identification improves the least by 20%.

9. Results: Computer Based
Recognition

10. Results: Reconstructing based on frequency
content
 Number of (10ms) frames in 10 seconds audio snippet = 667 frames.
 Number of possible orderings = 667! (intractable to break shredding by
bruteforce).
 Reconstructing by frequency content
 Greedly match the left and right edge of subsequent frames in frequency domain.
 Can reconstruct if audio is broken in 5 or less segments

11. Critique of work(1slide)
 Sound subsampling alone is not sufficient for privacy
preserving (at least for people counting, and gender
identification).
 Shredding can be attacked (As they mentioned at the
end of paper)
 Should compare against other methods (like filtering or
perturbing the speech frequency range in the audio
collected)