Luca Becker¹, Haitham Afifi², and Rainer Martin¹

¹Ruhr-Universität Bochum, ²Hasso-Plattner-Institut

In this demo, we showcase the utility of privacy-preserving acoustic feature extraction in ASN-based applications. For this purpose, we make use of features that are extracted from a deep neural network which is optimized via an adversarial training procedure. These features are further adopted for neural network-based gender recognition (trust model) and intercepted by a threat model that performs privacy-invasive speaker identification. We show that, in spite of the tight relation of both tasks, the features prevent speaker identification attacks while maintaining sufficient performance on the gender recognition task. The implementation of the feature extractor, trust model and threat model is based on the MARVELO framework on various interconnected Raspberry Pis. Predicated on this setup two different approaches are presented: firstly, we showcase successful speaker identification using standard features and secondly, we demonstrate how using privacy-preserving features prevents information leakage and consequently an exploitation by the attacker.