More security for Alexa, Siri & Co.


CRISP researchers are developing a prototype for privacy-protective speech recognition

In the Cybersecurity profile area of TU Darmstadt, researchers are working on various challenges in the area of Cybersecurity and Privacy. The topic of secure voice-controlled services is an example of cutting-edge research, as will be held in future in the new building on Kantplatz, for which the topping-out ceremony is celebrated today.

© Patrick Bal

Daily lives are ubiquitously influenced by Amazons "Alexa", Apples "Siri", Google Assistant or microsofts "Cortana". More than 2 billion smartphone users have immediate access to these services, meanwhile smart-home-devices like Amazon Echo, Apple HomePod or Google Home become more and more popular. This is not only true for the private sector, but is also in a trial state to boost productivity at work.

A continuous stream of audio recordings is transmitted to the cloud to enable speech recognition. This process poses considerable risks due to the contained biometric data and potentially sensitive and confidential data. Besides the risk of loosing valuable information (private and trade), there is also a potential danger of "Fake Recordings" where artifically generated recordings sound authentic and deliver a compromising content.

Software architecture „VoiceGuard“

Researchers of TU Darmstadt, under the leadership of Prof. Ahmad-Reza Sadeghi and Prof. Thomas Schneider, including speech recogintion expert Prof. Korbinian Riedhammer of the University of Applied Sciences Rosenheim have developed a new software architecture called „VoiceGuard“, to aggravate the potential security risks, as best as possible. VoiceGuard uses Intel Software Guard Extensions (SGX), to isolate speech processing of either the service provider or the users and provide security for all sensitive information. This enables personal privacy as well as safety of the intellectual property of the service provider. The evaluation of a first prototype shows that VoiceGuards privacy-guarding abilities enable speech recognition in real-time. Due to the generic architecture the concept is very adaptable to similar tasks, like emotion recognition. VoiceGuard will be presented at INTERSPEECH 2018, an inter­national top conference concerning language processing, in September 2018. 

Publications concerning "VoiceGuard"

Press release of TU Darmstadt

show all news