Week 2: Acoustic emanations
What a keyboard's sound reveals about what was typed on it, and how far the effect generalises
1 March 2027
A channel needs no contact and no electronics: a microphone in the room is enough, once a language model supplies what the signal alone cannot.
Every key on a mechanical or membrane keyboard produces a slightly different sound, because it strikes at a slightly different position and with a slightly different mechanism. Asonov and Agrawal’s 2004 paper “Keyboard Acoustic Emanations” showed that a simple recording of typing, fed through neural-network classification trained on that same keyboard, recovers a majority of keystrokes correctly — no bug in the keyboard, no malware on the machine, only a sound the room was always making.
The result generalises further than a single keyboard: later work extended it to recordings made over VoIP calls and to keyboards the classifier had never heard before, trained instead on the statistical regularities of the language being typed. The mechanism is acoustic and mechanical; the leak it produces is closer to a language-modelling problem than a signal-processing one, which is why the channel stays open even against keyboards designed to sound quieter.
The generalisation that matters, though, is not to another keyboard. Genkin, Shamir and Tromer’s 2014 paper “RSA Key Extraction via Low-Bandwidth Acoustic Cryptanalysis” recorded the machine rather than the thing typed on it: the capacitors and coils in a laptop’s voltage regulator vibrate as the processor’s load changes, and the resulting noise sits under 20 kHz — within range of a plain mobile phone placed beside the computer, or a more sensitive microphone four metres away. Different decryption keys produced measurably different acoustic signatures, enough to recover a full 4096-bit RSA key in about an hour. That upper limit is roughly five orders of magnitude below the clock driving the computation, which is the point rather than a caveat: the channel does not observe operations, it observes the envelope of very many of them, and the key falls out of the aggregate.
The defensive cost is the instructive part, because nobody shielded anything. The response shipped in GnuPG 1.4.16, alongside the disclosure, was exponentiation blinding — randomising the intermediate values so that a chosen ciphertext no longer produces a predictable profile to measure. An acoustic emanation from a power supply was closed by a change to arithmetic, which is week 11’s argument arriving nine weeks early.
Outline#
- what makes one key’s sound distinguishable from another’s, in mechanical terms: strike position, plate resonance, and the distance to the recorder
- why the decisive improvement came from the language being typed rather than from the microphone — the point at which this stopped being a signal-processing result
- the channel’s own limits: what a second, unseen keyboard does to a trained classifier, and why that limit is smaller than it sounds
- the same channel with no keyboard in it: acoustic emanation from a laptop’s own power supply, what a sub-20 kHz ceiling forces the analysis to do instead of reading operations, and why the fix was arithmetic rather than soundproofing
Readings
- Dmitri Asonov and Rakesh Agrawal (2004). Keyboard Acoustic Emanations. IEEE Symposium on Security and Privacy 2004, 3-11.
- Li Zhuang, Feng Zhou and J. D. Tygar (2005). Keyboard Acoustic Emanations Revisited. ACM CCS 2005, 373-382.
- Daniel Genkin, Adi Shamir and Eran Tromer (2014). RSA Key Extraction via Low-Bandwidth Acoustic Cryptanalysis. CRYPTO 2014, LNCS 8616, 444-461.
Teaching team
- Idris Fenn — tutor