Researchers Demonstrate Subliminal Smart Device Commands That Have Potential for Malicious Attacks
Researchers in the United States and China have been performing tests in an effort to demonstrate that "hidden" commands, or those undetectable to human ears, can reach AI assistants like Siri and force them to perform actions their owners never intended. The research was highlighted in a piece today by The New York Times, suggesting that these subliminal commands can dial phone numbers, open websites, and more potentially malicious actions if placed in the wrong hands.
A group of students from the University of California, Berkeley and Georgetown University published a research paper this month, stating that they could embed commands into music recordings or spoken text. When played near an Amazon Echo or Apple iPhone, a person would just hear the song or someone speaking, while Siri and Alexa "might hear an instruction to add something to your shopping list." Or, more dangerous, unlock doors, wire money from your bank, and purchase items online.
The method by which the students were able to accomplish the hidden commands shouldn't be a concern for the public at large, but one of the paper's authors, Nicholas Carlini, believes malicious parties could already be making inroads with similar technology.
“We wanted to see if we could make it even more stealthy,” said Nicholas Carlini, a fifth-year Ph.D. student in computer security at U.C. Berkeley and one of the paper’s authors.
Mr. Carlini added that while there was no evidence that these techniques have left the lab, it may only be a matter of time before someone starts exploiting them. “My assumption is that the malicious people already employ people to do what I do,” he said.
Last year, researchers based at Princeton University and Zheijiang University in China performed similar tests, demonstrating that AI assistants could be activated through frequencies not heard by humans. In a technique dubbed "DolphinAttack," the researchers built a transmitter to send the hidden command that dialed a specific phone number, while other tests took pictures and sent text messages. DolphinAttack is said to be limited in terms of range, however, since it "must be close to the receiving device."
DolphinAttack could inject covert voice commands at 7 state-of-the-art speech recognition systems (e.g., Siri, Alexa) to activate always-on system and achieve various attacks, which include activating Siri to initiate a FaceTime call on iPhone, activating Google Now to switch the phone to the airplane mode, and even manipulating the navigation system in an Audi automobile.
In yet another set of research, a group at the University of Illinois at Urbana-Champaign proved this range limitation could be increased, showing off commands received from 25 feet away. For the most recent group of researchers from Berkeley, Carlini told The New York Times that he was "confident" his team would soon be able to deliver successful commands "against any smart device system on the market." He said the group wants to prove to companies that this flaw is a potential problem, "and then hope that other people will say, 'O.K. this is possible, now let's try and fix it.'"
For security purposes, Apple is stringent with certain HomeKit-related Siri commands, locking them behind device passcodes whenever users have passcodes enabled. For example, if you want to unlock your front door with a connected smart lock, you can ask Siri to do so, but you'll have to enter your passcode on an iPhone or iPad after issuing the command. The HomePod, on the other hand, purposefully lacks this functionality.