Dolphin Attack on Speech Recognition Software

A very interesting new technique for commanding speech recognition devices has been successfully tested against several of these systems. Chinese researchers are calling it the DolphinAttack technique because it takes the voice commands and raises the frequency to a level that is undetectable by human ears. It was tested on seven different systems on 16 different devices with significant success. What does this mean? It means that someone could command our smartphones, our personal assistants, or our vehicles to do tasks without us realizing what is happening.

This is a scary prospect. While having your windshield wipers turn on randomly may be unsettling, it may not put you in too much danger. However, the researchers were able to successfully command a speech recognition system to open a back door. If that opened door is in your car or your home, it could be very serious indeed.

So, what can be done? The only real solutions for the time being are:

1) Don’t use these systems at all,

2) Turn them off in the device’s settings when not in use, or

3) Only use them where the risk is lower.

The advice to the manufacturers is to configure these devices to ignore any frequencies at 20 kHz or above.

It’s possible that patches may be released to address this issue and if they are, make sure to apply them right away. The devices tested included Siri, Google Now, Samsung S Voice, Cortana, and the speech recognition system in the Audi Q3.

There were some issues when these were tested. Background noise, for example, decreased the success rate by up to 70%. That doesn’t change the fact that this issue does exist and perhaps we need our dogs to be our copilots and listen for these frequencies until they are resolved.