Learning privacy-preserving representation of audio data with adversarial learning: The usage of adversarial learning to address privacy problems in smart audio processing devices

Abstract

Recently, the development of IoT leads to numerous automated machine listening systems be ing introduced. In the audio signals processed by these systems, human voice also exists, which poses a threat of leakage of privacy information. This thesis investigates one potential solution for the privacy problem in smart audio devices: learning a privacy-preserving representation of audio signal using an adversarial learning setup. The target machine listening task for such rep resentation is sound event classifcation, and the privacy criterion is that human speech can not be discriminated in the signal. Basic adversarial learning works as expected when the speech discriminator in the adversarial system cannot discriminate speech information; however, residual privacy information can still be recovered. Further improvements to the adversarial learning setup is needed for the audio representation to achieve privacy preservation.