Sound Shredding: Privacy Preserved Audio Sensing

Sound provides valuable information about a mobile user's activity and environment. With the increasing large market penetration of smart phones, recording sound from mobile phones' microphones and processing the sound information either on mobile devices or in the cloud opens a window to a large variety of mobile applications that are context-aware and behavior-aware. On the other hand, sound sensing has the potential risk of compromising users' privacy. Security attacks by malicious software running on smart phones can obtain in-band and out-of-band sound information to infer the content of users' conversation. In this paper, we propose two simple yet highly effective methods called {\em sound shredding} and {\em sound subsampling}. Sound shredding mutates the raw sound frames randomly just like paper shredding and sound subsampling randomly drops sound frames without storing them. The resulting mutated sound recording makes it difficult to recover the text content of the original sound recording, yet we show that some acoustic features are preserved which retains the accuracy of context recognition.