How we reverse engineered the Cuban “sonic weapon” attack
Examining overlooked clues reveals how ultrasound could have caused harm in Havana
Kevin Fu, Wenyuan Xu and Chen Yan, IEEE Spectrum
Throughout last year, mysterious ailments struck dozens of U.S. and Canadian diplomats and their families living in Cuba. Symptoms included dizziness, sleeplessness, headache, and hearing loss; many of the afflicted were in their homes or in hotel rooms when they heard intense, high pitched sounds shortly before falling ill. In February, neurologists who examined the diplomats concluded that the symptoms were consistent with concussion, but without any blunt trauma to the head. Suggested culprits included toxins, viruses, and a sonic weapon, but to date, no cause has been confirmed.
We found the last suggestion—a sonic weapon—intriguing, because around the same time that stories about health problems in Cuba began appearing, our labs, at the University of Michigan–Ann Arbor, and at Zhejiang University in China, were busy writing up our latest research on ultrasonic cybersecurity. We wondered, Could ultrasound be the culprit in Cuba?
On the face of it, it seems impossible. For one thing, ultrasonic frequencies—20 kilohertz or higher—are inaudible to humans, and yet the sounds heard by the diplomats were obviously audible. What’s more, those frequencies don’t propagate well through air and aren’t known to cause direct harm to people except under rarefied conditions. Acoustic experts dismissed the idea that ultrasound could be at fault.
Then, about six months ago, an editor from The Conversation sent us a link to a video from the Associated Press, reportedly recorded in Cuba during one of the attacks.
The editor asked us for our reaction. In the video, you can hear a piercing, metallic sound—it’s not pleasant. Watching the AP video frame by frame, we immediately noticed a few oddities. In one sequence, someone plays a sound file from one smartphone while a second smartphone records and plots the acoustic spectrum. So already the data are somewhat suspect because every microphone and every speaker introduces some distortion. Moreover, what humans hear isn’t necessarily the same as what a microphone picks up. Cleverly crafted sounds can lead to auditory illusions akin to optical illusions.
The AP video also includes a spectral plot of the recording—that’s basically a visual representation of the intensities of the various acoustic tones present, arranged by frequency. Looking closely, we noticed a spectral peak near 7 kilohertz and a dozen other less-intense tones that formed a regular pattern with peaks separated by approximately 180 hertz. What could have caused these ripples every 180 Hz? And what kind of mechanism could make an ultrasonic source produce audible sound?
As the questions began to mount, it still didn’t make sense to us, and that seemed like an excellent reason to dig deeper.
We also felt an obligation to investigate. Our own research had taught us that ultrasound can compromise the security of many types of sensors found widely in medical devices, autonomous vehicles, and the Internet of Things. For the last decade, two of us (Fu and Xu) have been collaborating on embedded security research, with the goal of discovering physics-based engineering principles and practices that will make automated computer systems secure by design. For example, Xu’s 2017 paper “DolphinAttack: Inaudible Voice Commands” describes how we used ultrasonic signals to inject inaudible voice commands into speech recognition systems such as Siri, Google Now, Samsung S Voice, Huawei HiVoice, Cortana, Alexa, and the navigation system of an Audi automobile.
The Cuban ultrasonic mystery was too close to our research to ignore.
One thing we knew going into this investigation is that acoustic interference can occur where you least expect it. Several years ago, Fu became annoyed by an ear-piercing sound coming from a lightbulb in his apartment. He took spectral measurements and noticed that the lightbulb tended to shriek when the air conditioner turned on. He eventually concluded that the compressor was pumping coolant through its pipes at the same resonant frequency of the filament in the bulb. Normally, this wouldn’t be a problem. But in this case, the coolant pipes ran through the ceiling and mechanically coupled to the ceiling joist supporting the lightbulb. The superintendent opened up the ceiling and separated the joist from the pipe with a piece of duct tape, to dampen the unwanted coupling. The sound stopped.
We also knew that ultrasound isn’t considered harmful to humans—for the most part. Misused, an ultrasonic emitter that’s in direct contact with a person’s body can heat tissues and damage organs. And the U.S. Occupational Safety and Health Administration (OSHA) warns that audible subharmonics caused by intense airborne ultrasonic tones can be harmful. Thus, U.S. standards on ultrasonic emissions build in safety margins to account for those subharmonics. The Canadian government, meanwhile, has ruled that humans can be directly harmed by airborne ultrasound at sound pressures of 155 decibels or higher—which is louder than a jet taking off at 25 meters. That ruling also notes that “a number of ‘subjective’ effects have been reportedly caused by airborne ultrasound, including fatigue, headache, nausea, tinnitus and disturbance of neuromuscular coordination.”
Of course, even at 155 dB, ultrasonic tones remain inaudible. Unless they’re not—more on this in a bit.
To make the problem tractable, we began by assuming that the source of the audible sounds in Cuba was indeed ultrasonic. Reviewing the OSHA guidance, Fu theorized that the sound came from the audible subharmonics of inaudible ultrasound. In contrast to harmonics, which are produced at integer multiples of a sound’s fundamental frequency, subharmonics are produced at integer divisors (or submultiples) of the fundamental frequency, such as 1/2 or 1/3. For instance, the second subharmonic of an ultrasonic 20-kHz tone is a clearly audible 10 kHz. Subharmonics didn’t quite explain the AP video, though: In the video, the spectral plot indicates tones evenly spaced every 180 Hz, whereas subharmonics would have appeared at progressively smaller fractions of the original frequency. Such a plot would not have the constant 180-Hz spacing.
Fu explained his theory to Chen Yan, a Ph.D. student in Xu’s lab. Yan wrote back: It’s not subharmonics—it’s intermodulation distortion.
Intermodulation distortion (IMD) is a bizarre effect. When multiple tones of different frequencies travel through air, IMD can produce several by-products at other frequencies. In particular, second-order IMD by-products will appear at the difference or the sum of the two tones’ frequencies. So if you start with a 25-kHz signal and a 32-kHz signal, the result could be a 7-kHz tone or a 57-kHz tone. These by-products can be significantly lower in frequency while maintaining much of the intensity of the original tones.
IMD is well known to radio engineers, who consider it undesirable for radio communication. The sounds don’t have to travel through air; any “nonlinear medium” will do. A medium is considered nonlinear if a change in the output signal is not proportional to the change in the input. Acoustic devices such as microphones and amplifiers can also exhibit nonlinearity. One way to test for it is to send two pure tones into an amplifier or microphone and then measure the output. If additional tones appear in the output, then you know the device is nonlinear.
Computer science researchers have explored the physics of IMD. In the DolphinAttack paper, we used ultrasonic signals to trick a smartphone’s voice-recognition assistant. Because of nonlinearity in the smartphone’s microphone, the ultrasound produced by-products at audible frequencies inside the circuitry of the microphone. Thus, the IMD signal remains inaudible to humans, but the smartphone hears voices. In an early 2017 paper, Nirupam Roy, Haitham Hassanieh, and Romit Roy Choudhury at the University of Illinois at Urbana-Champaign described their BackDoor system [PDF] for using ultrasound and IMD to jam spy microphones, watermark music played at live concerts, and otherwise create “shadow” sounds.
Some composers and musicians have also used IMD to create synthetic sounds, combining audible tones to create other subliminal, audible tones. For example, in their 1987 book The Musician’s Guide to Acoustics, Murray Campbell and Clive Greated note that the last movement of Jean Sibelius’s Symphony No. 1 in E minor contains tones that lead to a rumbling IMD. The human ear processes sound in a nonlinear fashion, and so it can be “tricked” into hearing tones that weren’t produced by the instruments and that aren’t in the sheet music; those subliminal tones are produced when the played tones combine nonlinearly in the inner ear.
Back to our quest: Knowing that intermodulation distortion between multiple ultrasonic signals can cause lower-frequency by-products, we next set about simulating the effect in the lab, aiming to replicate what we observed in the AP News video. We used two signals: a pure 25-kHz tone and a 32-KHz carrier tone that had its amplitude modulated by a 180-Hz tone. (Our technical report, “On Cuba, Diplomats, Ultrasound, and Intermodulation Distortion” [PDF], goes into more detail on the math of how we did this.) The result was clear: Strong tones appeared at 7 kHz with repeating ripples separated by 180 Hz.
We then followed up with live experiments. As in the simulation, we used two ultrasonic speakers to emit the signals, one as a 180-Hz sine wave amplitude modulated over a 32-kHz carrier, and the second as a single-tone 25-kHz sine wave. We used a smartphone to record the result. IMD caused by the air and the smartphone microphone created the telltale 7-kHz signal. This video shows the experimental setup:
If you look closely at the spectral plot displayed on the smartphone, you’ll notice some higher-order IMD by-products, at 4 kHz and beyond, as well as several other frequencies. Interestingly, although we could hear the 7-kHz tones during the experiment, we couldn’t hear the 4-kHz tones recorded by the smartphone. We suspect that the 4-kHz tones partly resulted from secondary IMD within the microphone itself. In other words, the microphone was hearing an acoustic illusion that we couldn’t hear.
For fun, we also experimented with using an ultrasonic carrier to eavesdrop on a room. In this kind of setup, a spy places a microphone to pick up speech and then uses the relatively low-frequency audio signal to modulate the amplitude of the carrier wave. The carrier wave then gets picked up by an ultrasonic-capable sensor located some distance away and demodulated to recover the original audio. In our experiments, we selected a song to stand in for the audio signal recorded by an eavesdropping microphone: Rick Astley’s 1980s hit “Never Gonna Give You Up.” We amplitude modulated the song on a 32-kHz ultrasonic carrier. When we introduced a 25-kHz sine wave to interfere with this covert ultrasonic channel, IMD in the air produced a 7-kHz audible tone with ripples associated with the tones of the song, which was then picked up by the recording device. The computer played the song after software demodulation.
One thing to note in the video is that the metallic sounds near 7 kHz are audible only at the point where the two signals cross. When the signals do not intersect, you can’t hear the 7-kHz tone, but the demodulator can still play the covert song. That finding is consistent with what some diplomats reported in Cuba: The sounds they heard tended to be confined to a part of the room. When they moved just a few steps away, the sound stopped.
So if the sources of the sound in Cuba were ultrasonic, what could they have been? There are many sources of ultrasound in the modern world. At Michigan, our offices are bathed in 25-kHz signals coming from ceiling-mounted ultrasonic room-occupancy sensors. We’ve removed the devices closest to our lab equipment, but just last month we discovered a new one. [To learn more about our travails with these sensors, see “How an Ultrasonic Sensor Nearly Derailed a Ph.D. Thesis.”] Another source is ultrasonic pest repellents against rodents and insects. (This blog post describes a family’s encounter with such a device in the Havana airport.) And some automobiles contain ultrasonic emitters.
While the equipment we used in our Cuban re-creation is relatively bulky, ultrasonic emitters can be quite tiny, no larger than a piece of Rolo candy. Online, we found a manufacturer in Russia that sells a fashionable leather clutch that conceals an ultrasonic emitter, presumably to jam recording devices at cocktail parties. We also found electronics stores that carry high-power ultrasonic jammers that cause microphones to malfunction. One advertised jammer emits 120-dB ultrasonic interference at a distance of 1 meter. That’s like standing next to a chainsaw. If a signal from that caliber jammer were to combine with a second ultrasonic source, audible by-products could result.
While the math leads us to believe that intermodulation distortion is a likely culprit in the Cuban case, we haven’t ruled out other null hypotheses that may account for the discomfort that diplomats felt. For example, maybe the tones people heard didn’t cause their symptoms but were just another symptom, a clue to the real cause. Or maybe the sounds had some sort of nonauditory effect on people’s hearing and physiology, through bone conduction or some other known phenomenon. Microwave radiation is another theory. One positive outcome from all this would be if more computer scientists were to master embedded security, signal processing, and systems engineering.
Even if our hypothesis is correct, we may never learn the definitive story. The parties responsible for the ultrasonic emitters would have already figured out by now that their devices are to blame and would have removed or deactivated them. But whether our hypothesis is correct or not, one thing is clear: Ultrasonic emitters can produce audible by-products that could have unintentionally harmed diplomats. That is, bad engineering may be a more likely culprit than a sonic weapon.
About the Authors
Kevin Fu is a Fellow of the IEEE and an associate professor of computer science and engineering at the University of Michigan–Ann Arbor, where he leads the Security and Privacy Research Group. He’s also chief scientist of the health-care cybersecurity startup Virta Labs. Wenyuan Xu is professor and chair of the department of systems science and engineering at Zhejiang University. Xu’s Ubiquitous System Security Lab (USSLab) has twice been recognized by the Tesla Security Researcher Hall of Fame. Chen Yan is a Ph.D. student at Zhejiang University.
To Probe Further
The authors’ technical report “On Cuba, Diplomats, Ultrasound, and Intermodulation Distortion” [PDF] (Technical Report CSE-TR-001-18, University of Michigan, Computer Science & Engineering, 1 March 2018) provides additional details on their simulation and experiments to reverse engineer the Cuban embassy “sonic weapon.”
AP News’s Josh Lederman and Michael Weissenstein were the first to report the Cuban sound recording, in “Dangerous sound? What Americans heard in Cuba attacks,” 13 October 2017.
For more on how sounds can be synthesized using intermodulation distortion, see “Sound Synthesis and Auditory Distortion Products,” by Gary S. Kendall, Christopher Haworth, and Rodrigo F. Cádiz, in Computer Music Journal, 38(4), MIT Press, Winter 2014.
A number of people have suggested that microwaves, rather than ultrasound, may have been at work in Cuba. See, for example, James C. Lin’s article “Strange Reports of Weaponized Sound in Cuba,” in IEEE Microwave Magazine, January/February 2018, pp. 18-19. A remaining question is whether microwaves could have produced the high-pitched sounds recorded by the smartphone in the AP News video.
IN ENGLISH LANGUAGE
FOR PEOPLE WHO READ IN ENGLISH: ORIGINAL DOCUMENTS IN ENGLISH OR TRANSLATED. PUBLICATION DOES NOT MEAN WE ENDORSE OR REJECT CONCLUSIONS OR STATEMENTS OF AUTHORS