Taking Away the Beep: Timing-Tag Removal as a Defense Against Covered-Hand PIN Attacks

Covering your hand while typing a PIN at an ATM or card terminal does not stop a camera from learning the PIN. Recent attacks train a video model on the covering hand's small, involuntary movements and recover the digits underneath. They get the right 5-digit PIN within three guesses 30% of the time on ATMs and 53.8% of the time on tilted card terminals. These attacks share a quiet dependency. Before the model can classify a keystroke, it must know when the keystroke happened, and it learns that from the terminal itself: the ATM's beep or the asterisk that appears on the card terminal screen. Rare-event physics has a name for this structure. A search for a faint signal succeeds or fails largely at the trigger, the step that decides which moments count as candidate events, and a trigger built on a coincidence between two channels is only as strong as the true events are distinguishable from accidental ones.

This note proposes treating the terminal's feedback as the attacker's trigger and disrupting it. Real keystrokes would stay confirmed to the user through a channel the camera cannot see, such as the tactile click under the finger or a narrow-angle display. Everything the camera can hear or see would be delayed by random amounts or mixed with decoy events. The first question is whether the attacker even needs the tag. The attack papers themselves call tag-free segmentation of keystrokes from video alone an open problem. The second question is how much of the attack survives when the tag becomes noisy rather than absent. The answers decide whether the right defense is simply "turn off the beep" or something more careful.

Core Claim

Covered-hand PIN attacks are two-stage systems: first cut the video into keystrokes, then classify each one. The published accuracy figures measure both stages running with a perfect external timing tag. We claim three things. First, a large, measurable share of that accuracy comes from the tag rather than from the hand kinematics. Second, removing the tag outright is necessary but may not be sufficient, because an adaptive attacker can learn to find keystrokes from the moments when the hand slows down. Third, feedback that is decorrelated from true press times, with random delays plus decoy events matched to the attacker's own segmentation features, can push an adaptive attacker toward chance at an acceptable usability cost.

The contribution from physics is a way of thinking about the attack and a way of scoring it, not a new detector. We describe the attacker as a coincidence search: an audio or screen event must line up with a visible hand deceleration. We then report the defense the way rare-event experiments report a background cut, as the pair (fraction of true keystrokes the attacker still segments correctly, number of false segments per true keystroke). A single accuracy number hides where the attack fails.

Grounding in Prior Work

The motivating evidence comes from the attack literature. Hand Me Your PIN! trains a network that pairs per-frame image features with a sequence model (a CNN-LSTM) on 5,800 covered-hand entries from 58 people on real metal ATM keypads. It uses the buzzer beep to segment keystrokes and names visual-only segmentation as open. Your PIN is Mine does the same on card terminals with 142 participants and 13,850 PINs, using the on-screen asterisks for timing. It finds tilted mounts most vulnerable and the attack robust to 2 m. Both papers suggest only heuristic countermeasures, such as physically compact PINs or screen flicker. Neither evaluates a defense against an attacker that is retrained on the defended data, which is what we do. PrivacyScout adds a constraint on design: PIN entry stayed vulnerable at every distance and angle tested because observers track finger motion. Any defense that works on the display alone is therefore ruled out, and our defense works on the timing channel instead. Heat of the Moment shows that residual heat on plastic keypads leaks PINs after entry. We leave this out of scope and say so, because timing decoys do nothing against it.

The method comes from rare-event physics, where true events are defined by coincidences in time. SNO+ picks out solar neutrino interactions on carbon-13 by requiring a delayed coincidence with the subsequent nitrogen-13 decay (First Evidence of Solar Neutrino Interactions on 13C). PandaX-4T pairs a prompt light flash with a delayed charge signal to locate and classify events (PandaX-4T 1.54 tonne-year results). The covered-hand attacker does the same thing with a beep and a hand motion. The reporting discipline also comes from physics. α-event rejection in point-contact germanium detectors reports both near-complete surface-event rejection and 99.8% retention of true events, including where rejection degrades near the central contact. PandaX-III topological discrimination reports roughly seven orders of magnitude of background suppression at above 20% signal efficiency. The lesson we import is that a cut is always described by two numbers. The same PandaX-III paper also sets a limit on our expectations: when its detector lacks an absolute start-time tag, the fallback depth estimate from how much the electron cloud spreads keeps only 75% of signal for a modest background reduction. Losing the timing tag hurts, but a physical fallback variable still carries information. That is exactly the situation we expect the attacker to be in once the beep is gone.

Research Program

The work has three parts, and the later parts depend on what the earlier ones find.

Part one measures what the tag is worth to the attacker. We rebuild the published CNN-LSTM attackers, run with the beep or asterisk (tag-aided), and then build the strongest tag-free attacker we can. That attacker has a segmentation network that proposes keystroke times from hand, wrist and forearm motion, feeding the same keystroke classifier, and is trained end to end with a loss that sums over possible alignments between video and digits. We call the difference in Top-3 PIN accuracy between the tag-aided and tag-free attackers the tag's value. If the tag-free attacker falls to near the ~8% human baseline, the simple defense wins and the rest of the program shrinks to a usability study of silent feedback.

Part two builds decorrelated feedback. Real presses produce feedback after a random delay drawn from a window $J$. $J$ is bounded above by feedback latency users tolerate and below by keeping feedback events in their correct order: $J$ must stay under the typical time between keystrokes. On top of this we emit decoy feedback events as a random (Poisson) process with rate $r_d$. The key design principle from physics is that background passes a cut only when it looks like signal in the discriminating variables. So decoys are timed to moments when the hand visibly slows down or pauses, which the terminal can detect from its own keypad timing and a cheap internal motion model, rather than scattered uniformly in time. The user never relies on the public channel to count digits. Confirmation comes through tactile key travel and, on card terminals, a hooded or narrow-angle digit counter.

For a naive attacker that accepts any feedback event within a window $\tau$ of a candidate hand motion, the standard result for random, uncorrelated event rates gives the expected number of false segments per true keystroke:

$$ N_{\text{false}} \approx 2\,\tau\, r_d , $$

where $\tau$ must itself grow roughly with $J$ for the attacker to keep its true-keystroke efficiency. We use this only as a first-order design guide. Decoys placed at hand pauses are meant to beat it, because an adaptive attacker cannot reject them by motion features alone.

Part three tests the defense against an attacker that knows it. The attacker is retrained on defended recordings with the same compute and tuning budget as the undefended attacker. We then run a small in-lab usability study of the feedback schemes.

Baselines

The naive reference point is the undefended terminal attacked by the published models. For ATMs, Hand Me Your PIN! reports Top-3 accuracy of 30% for 5-digit PINs and 41% for 4-digit PINs on a known keypad, against about 8% for human observers. For card terminals, Your PIN is Mine reports 53.80% Top-3 for 5-digit PINs on the iPP 320 at a 45° tilt. We first reproduce these numbers on the same data splits. Our reimplementation must come within a few points of them before any defense result is trusted. We also report the chance floor for three guesses over the PIN space, which is 3 in 10,000 for 4-digit PINs.

The strongest existing attack is the tag-aided CNN-LSTM itself, and the strongest attacker on the defended question is our tag-free, end-to-end-segmenting variant. We treat that variant as a baseline we are obliged to make strong. It gets the same backbone, the same training data volume, and a hyperparameter search matched in trials to the undefended attacker. Any defense is scored against whichever attacker, tag-aided or tag-free, does better on it.

The strongest simple defense is silence: remove the beep or asterisk entirely. It is the baseline our proposal must beat. Beating it means that, against the adaptive attacker, decorrelated feedback yields meaningfully lower Top-3 accuracy than silence, or equal accuracy with better user confirmation, measured as entry errors and self-reported confidence.

Two ablations isolate the key ingredient. Jitter-only varies $J$ with no decoys. Uniform decoys places decoys at random times rather than at hand pauses. If feature-matched decoys do no better than uniform ones against the adaptive attacker, then the physics-derived design principle is not contributing and we will say so.

Why not just…?

Why not just turn off the beep and the asterisks? This is the most serious competitor, and it may be enough. The attack papers depend on these tags, and neither shows that visual-only segmentation works. Part one exists to settle this. If the tag-free attacker collapses toward the human baseline, silence is the answer, and our contribution becomes the measurement that justifies it. We expect it not to collapse fully. Hand deceleration at each press is a strong visual cue, and the PandaX-III experience suggests that a lost timing tag degrades rather than destroys discrimination. Silence also removes the user's audible confirmation, which the usability study will price.

Why not just shield better or choose physically compact PINs? Shielding is the defense these attacks already defeat: 53.80% Top-3 with the hand covered on tilted card terminals (Your PIN is Mine). Compact PINs shrink the effective PIN space the attacker must search and rely on user compliance. Our defense is set by the terminal and needs no change in user behavior.

Why not just scramble the keypad layout? The ATM attack targets physical metal keypads, where the layout cannot change. On touchscreen card terminals, scrambling breaks motor memory, which is the consistency that touch-dynamics verification relies on (MobileTouchDB). It also fails when the camera catches the screen across sessions, as the 23–46% iterative-attack success in GazeTouchPass shows for split-channel schemes.

Why not just switch to gaze input? Gaze input nearly eliminates video attacks, with 0.05% success in CueAuth. It costs 26.35 s per entry against 3.73 s for touch, and entry success falls to 82.72%, which is not viable at a checkout. Our target is to keep touch-level entry time.

Proposed Experiments

The data are the two covered-hand corpora from the attack papers, requested from their authors. If they are unavailable, we re-collect under their published protocols: the same keypad models, tilt and distance conditions, and participant-disjoint splits. Because keystroke times are known in these recordings, defenses on the audio channel can be simulated after the fact by replacing the soundtrack with jittered or decoy beeps. On card terminals, the asterisk region is masked and re-rendered. This lets us sweep $J$ and $r_d$ without new data collection, and we then confirm the best settings with a live recording of about 30 participants.

Metrics are Top-3 PIN accuracy and per-digit Top-3 accuracy, as in the source papers. We add the physics-style pair of keystroke segmentation efficiency and false segments per keystroke, and plot defense curves of attacker accuracy against added latency. Usability is measured by entry time, entry error rate, and NASA TLX workload, a standard questionnaire. We compare against the CueAuth touch figures of 3.73 s and 93.38% success as a reference for acceptable entry.

The decisive checks are three. First, the tag's value from part one: if it is small, the proposal reduces to recommending silence. Second, whether decorrelated feedback beats silence against the adaptive attacker at matched budget. Third, whether decoys matched to hand pauses beat uniformly timed decoys. A finding that silence already reaches near chance, or that decoys give no gain over jitter, would be reported as the result rather than argued around.