Papers recover keys, printouts, and an RSA key from sound, and name no secret successor. A 2004 IBM paper trains a network on key clicks. A 2005 Berkeley paper reconstructs English from ten minutes of typing with no labeled sample. A 2010 paper reads a dot-matrix printer from a nearby microphone. Two write-ups of one RSA attack, dated 2013 and 2014, do not print the same distance. A 2017 paper follows the clicks through a voice call. None of them names a still-secret program that replaced these demonstrations.
Key clicks, with and without a labeled sample
Dmitri Asonov and Rakesh Agrawal, at IBM Almaden, write in their 2004 symposium paper that PC keyboards, notebook keyboards, telephone pads, and ATM pads are vulnerable to an attack that tells keys apart by sound. A neural network recognizes the key. Besides a computer, they say, the hardware is a parabolic microphone, and the sound can be recorded from a substantial distance without opening the machine. Most PC tests used an IBM keyboard, serial 0953260. On a 300-click test, the correct key received the network's largest value in 79 percent of clicks. The second-largest value was right in 7 percent, the third in 2 percent, and the correct key was outside the top three in 12 percent. They report that typing style has little effect, so a network trained on one person can be used on another person at the same keyboard, and that a network trained on one keyboard can attack another of the same type with a drop in recognition.
Asonov and Agrawal, Keyboard Acoustic Emanations, 2004: https://agrawal-family.com/rakesh/papers/ssp04kba.pdf
Li Zhuang, Feng Zhou, and J. D. Tygar, in the proceedings of the ACM conference in Alexandria, November 2005, take a ten-minute recording of English typing and recover up to 96 percent of the characters, with no labeled training recording. They write that the recognizer built that way can then attack random text: in their experiments, 90 percent of five-character random passwords made only of letters can be generated in fewer than 20 attempts, and 80 percent of ten-character passwords in fewer than 75 attempts. They describe Asonov and Agrawal as reaching about 80 percent only when the network was trained on labeled sound from the same keyboard and the same person. The 2004 paper's own 79 percent figure is close to that "approximately 80 percent," and the conditions are not the same claim: one paper stresses a labeled sample, the later paper says it does not need one. Both papers are about keyboard sound. Neither names a government system.
Zhuang, Zhou, and Tygar, Keyboard Acoustic Emanations Revisited, November 2005: https://people.eecs.berkeley.edu/~tygar/papers/Keyboard_Acoustic_Emanations_Revisited/ccs.pdf
A printer, then a laptop tone at two printed distances
Michael Backes and coauthors, in a 2010 USENIX paper, recover English text from the sound of a dot-matrix printer when the microphone is close enough. In their experiments the attack recovers up to 72 percent of printed words, and up to 95 percent if contextual knowledge of the text is assumed, with the microphone 10 centimeters from the printer. After a training phase they describe the attack as automated. They also say they mounted it in the field, with privacy protections, in a doctor's practice, to recover medical prescriptions. A survey they cite says about 60 percent of doctors in Germany then used dot-matrix printers, about 190,000 doctors and more than 2.4 million records and prescriptions a day on average, and about 30 percent of banks, about 14,000 branches and more than 1.2 million documents a day. The paper does not name a secret collection program.
Backes and others, Acoustic Side-Channel Attacks on Printers, 2010: https://www.usenix.org/legacy/event/sec10/tech/full_papers/Backes.pdf
Daniel Genkin, Adi Shamir, and Eran Tromer, in a paper dated 18 December 2013, say many computers emit a high-pitched noise from vibrating electronic parts, and that the noise can leak a cryptographic key. They describe an attack on GnuPG's RSA that extracts full 4096-bit decryption keys from laptop computers of various models, within an hour, from the sound of decrypting chosen ciphertexts. They write that they demonstrated it with a plain mobile phone placed next to the computer, or with a more sensitive microphone placed 4 meters away. The same paper says a similar low-bandwidth attack can use the electric potential of the chassis. The acoustic claim in this write-up is the phone and the 4 meters.
Genkin, Shamir, and Tromer, 18 December 2013: https://cs-people.bu.edu/tromer/papers/acoustic-20131218.pdf
The version in the CRYPTO 2014 proceedings states the same target, full 4096-bit RSA keys from GnuPG within an hour, and then a different distance: a plain cellular phone next to the computer, or a sensitive microphone from 10 meters. It says the attacks apply to GnuPG 1.x through 1.4.15, released 15 October 2013, including the mitigation added in 1.4.14, which the authors say helps the attack. It says that after they disclosed the attack, new versions of GnuPG 1.x and of libgcrypt added countermeasures that are effective against the key-extraction attack, though not against acoustic key-distinguishing, with that distinction discussed in the extended version. Four meters in the December 2013 paper and ten meters in the 2014 proceedings are both printed. Neither paper fills the gap with an unnamed agency system. The phone, the microphone, GnuPG, and the stated distances are the record.
CRYPTO 2014 proceedings: https://iacr.org/archive/crypto2014/86160149/86160149.pdf
Clicks carried on a call
Alberto Compagno, Daniele Lain, Mauro Conti, and Gene Tsudik, in an arXiv text dated 11 March 2017 and marked to appear at ASIACCS 2017, describe keyboard sound picked up inside a voice-over-IP call. They call the attack Skype & Type. They write that Skype conveys enough audio to reconstruct keystrokes typed on the remote keyboard. Given some knowledge of the victim's typing style and keyboard model, they report a top-five accuracy of 91.7 percent in guessing a random key. The abstract also says the attack avoids prior assumptions of physical proximity and of precise profiling, and then states the 91.7 percent figure under "some knowledge" of style and model. It says the method still works with bandwidth fluctuation and with voice over the keystrokes, and that it applies to other software, naming Google Hangouts. The paper cites earlier keyboard-sound work. It does not name a classified successor to that work, or to the RSA attack, or to the printer attack.
Compagno, Lain, Conti, and Tsudik, arXiv:1609.09359v2, 11 March 2017: https://arxiv.org/pdf/1609.09359
What is named, and what is not
The devices in these papers are keyboards, telephone and ATM pads, a parabolic microphone, a dot-matrix printer, laptops, a mobile phone used as a microphone, and the microphones inside Skype and Hangouts. The software names are a neural-network recognizer, GnuPG and libgcrypt, Skype, and Hangouts. The distances and rates do not all match: 79 percent against an "approximately 80 percent," 4 meters against 10 meters, a labeled sample against none. No paper in this set prints a still-secret program that took over when these demonstrations were published.
