2. Machines perceive sound frequencies that humans cannot hear

[Rakhirandiseo]

The technique of adversarial attacks is the subject of research at the University of California. The study conducted there describes a scheme for fooling speech recognition systems based on machine learning. The researchers proposed introducing small changes to audio recordings of a voice that are 99.9% identical to the original recording, but completely change its content. Moreover, the technique they developed involves the use of hidden commands.

So far, the scheme created by the scientists is only iceland mobile database to certain devices and under certain conditions, but they believe that massive attacks using fake images and hidden commands are just around the corner. They have not yet managed to hack voice assistants “over the air,” but it is possible that they will succeed in the future.


Commands hidden inside audio files are not the only means of manipulating voice assistants. In 2017, a group of researchers from Zhejiang University invented an attack technique using ultrasound — they are inaudible to humans, but are keenly detected by AI. As it turns out, using sound commands, you can force Siri to make calls or perform other uncontrollable actions.

The authors named the new type of vulnerability DolphinAttack, but its implementation will require several parts, which, however, can be purchased at any radio store. The device assembled by scientists can control a "smart" speaker using ultrasound with a frequency of over 20 kHz, which the microphone software ignores, but technically continues to receive.