Ultrasonic attacks on Speech recognition systems

Researchers out of Zhejiang University have done ultrasound experiments on Speech recognition systems such as Siri, Google now, Amazon’s Alexia, and other Speech recognition devices to see if they can get modulated voice commands to work in frequencies of 20KHz or higher. The University has named the method the DolphinAttack.

The activation commands that they tried were things like “OK Google”, “Alexa”, and “Hey Siri”. The commands they gave after activation were things like “Open dolphinattack.com” and “Call 1234567890”. The experiment was tested on 7 speech recognition systems on 16 different devices. The attack was successful on all systems and all devices from various distances.

Some of the things that affected the attack were what the command was, such as “Call 1234567890” had much better results than “open dolphinattack.com”. Another thing is the distance from the device that was being attacked. The furthest attack was on Siri on the Iphone 4 and on Amazon’s Alexa on echo. Those two devices registered commands from over 6 feet away. The third thing that effected the attack was background noise. On the street the command “turn on airplane mode” was only successful only about 30% of the time where in a cafe it worked about 80% of the time and in a quite office it worked 100% of the time. The one thing that didn’t seem to affect the attack was what language the command was spoken in.

Some of the proposed defenses against this sort of attack were both hardware and software based. The reseachers suggest that the mics in the phone be enhanced by suppressing acoustic signals in the ultrasound range. The software based defense they propose is that it looks at the features of real voice commands vs modulated commands which have distinct acoustic features.

Sources: