Towards Opaque Audio Features for Privacy in Acoustic Sensor Networks

Inhalt:
In this paper we define a general scenario and develop first steps towards a framework for assessing the privacy and utility of audio features in acoustic sensor networks. We propose to use scalable feature sets which we derive from the cepstral modulation spectrum and introduce the notion of opaque features. Opaque features achieve a certain level of performance in clustering and classification tasks while they reveal a limited amount of information about audio signals. The proposed feature set offers a multitude of possibilities for balancing the performance in classification experiments (feature utility) and the amount of information disclosed. The utility of these features is measured via the Fisher discriminant whereas privacy is assessed via the mutual information of the feature vector and a highresolution representation of the audio signal. We show that the amount of information revealed can be successfully limited by a feature selection and temporal aggregation process.