Description
Speaker recognition is becoming more important as virtual assistants such as Cortana and Siri and devices such as Amazon Echo and Google Home are used on a daily basis.
Audio adversarial examples can be used to spoof the speaker recognition of these devices:
Audio data of speakers that could easily be classified by humans are misclassified by the recognition model.
Since the devices are often connected to other devices, private mail accounts, private cloud storage etc., adversarial examples impose a risk on security and privacy.
Building on previous research I show to what degree offsets influence the performance of speaker recognition for benign audio data as well as for adversarial audio examples.
Analyzing the effect of offsets for speaker recognition systems is relevant, since the security of such systems could potentially be compromised if an attacker exploits the effects that offsets have on the systems.
I also demonstrate one way an attacker can use the knowledge about these effects to make attacks more successful.
|
Thesis topic in co-operation with the Fraunhofer Institute for Applied and Integrated Security AISEC, Garching