Personal Voice Assistant Wake Word Jamming

Personal Voice Assistants (PVAs) such as Apple's Siri, Amazon's Alexa and Google Home are now ubiquitous. These devices continuously listen for a wake word that users speak to start interaction with the device. If this wake word recognition is disrupted, the device is not usable anymore. The wake word detection can be impeded by acoustic interference. Interference might be noise (e.g. background music, chatter, engine sounds) or a deliberate acoustic jamming signal. While wake word recognition algorithms might be designed with resilience against noise in mind they are usually not prepared to handle an attacker using a jamming signal. This work provides the first detailed study of acoustic Denial of Service (DoS) jamming attacks on the PVA wake word detection. We describe how a jamming signal should be designed such that wake word detection is jeopardised effectively while minimising jamming effort and jamming detectability. We study the impact of various noise features such as signal type, strength, timing, duration, frequency and bandwidth. Our work shows that accurately timed signals with a very short duration of only 2ms can prevent PVA operations reliably.