Voice assistants can detect typing on nearby devices, which could potentially be used to work out what a person is writing on their phone from up to half a metre away.
Ilia Shumailov at the University of Cambridge and his colleagues built a machine-learning system that could recognise the sound of tapping on a touchscreen and combined it with other artificial intelligence tools to try to determine what people were typing.
Shumailov and his team asked three volunteers to type randomly displayed 5-digit PIN or English words on a touchscreen device while audio was recorded by a microphone on a separate device nearby. The researchers then used the AI to try to figure out what the person had written.
Advertisement
The accuracy per character on the first guess for a 5-digit PIN ranged from 28 to 47 per cent when the person typing was 20 centimetres away from the recording device. But accuracy climbed to between 60 and 76 per cent with three guesses.
The accuracy fell as the distance between the person typing and the recording device increased 鈥� accuracy from 50 centimetres was approximately 20 per cent per character.
鈥淩ight now, it鈥檚 unlikely that people would use our attack. However, the world changes quickly and sensors only get better,鈥� says Shumailov. 鈥淭he fact that it鈥檚 possible is already very spooky.鈥�
鈥淭he implications reconfirm that having always-on cameras and microphones in our home will eventually come with privacy and security risks,鈥� says Hamed Haddadi at Imperial College London. 鈥淲hile this set-up is not easily possible for a third-party developer, it might just be possible for the voice assistant providers.鈥�
Companies that make these systems should be responsible for ensuring that such attacks aren鈥檛 possible, says Haddadi.
But the best solution is to avoid having microphones at home at all, says Shumailov. 鈥淚f that is unavoidable, it is best not to do anything privacy-sensitive next to microphones.鈥�
Reference:
