How scoring works¶
Every desktop session is scored live, a few seconds after its batches arrive, from typing (keys in code mode) and
from mouse movement:
- Keystrokes are cut into windows of 50 keys within bursts of typing (a pause over 10 s starts a new burst); windows that are mostly navigation keys are skipped.
- The first 3 windows form the session's template: by assumption the person who just signed in. Once a user has at
least 24 windows from earlier days, half of the template comes from their last 32 earlier-day windows
(
template: "history"). - Each later window is embedded by the keystroke model, compared with the template, and turned into a calibrated
log-likelihood ratio (
last_llr): positive means same person, negative different. - Mouse movement is cut into windows of 32 strokes (a pause over 300 ms ends a stroke), embedded by the mouse model and scored the same way against the session's first 3 mouse windows.
- One CUSUM adds up the evidence of both, in time order, and alarms when it crosses a threshold
(
detector: "fused"). Sessions with no typing template yet are scored on mouse alone (detector: "mouse"). Two thresholds run side by side:strict(target at most 1 false alarm per 100 hours of use) andsensitive(at most 10).
Measured results¶
| Measured | strict | sensitive |
|---|---|---|
| Takeovers detected within 3 min of active use | 72 % | 91 % |
| Median time to detection | 50 to 59 s | 25 to 40 s |
| False alarms per 100 h of genuine use | 3.4 | 13 |
| Mouse only: takeovers detected within 3 min | 18 % | 27 % |
Fused detector, cross-fitted on held-out users: typing from Clarkson II paired with mouse sessions of other held-out people (BB-MAS, Chao Shen, Markov CAD and others), as if the same person typed and moused at once. That pairing assumes both are in use together and independent, so treat these as an upper bound until they are measured on real sessions with both. Mouse alone is weak: it mainly covers sessions without typing.
| Measured | strict | sensitive |
|---|---|---|
| Takeovers detected within 3 min of the intruder's typing | 64 % [58, 71] | 89 % [85, 94] |
| Median time to detection | 136 s | 61 s |
| False alarms per 100 h of genuine typing | 0.8 (upper bound 4.5) | 16 |
Keystroke detector on held-out users of the Clarkson II desktop dataset (23 victims, one model seed), template with history, thresholds set on other users than they were tested on. Times count active typing, not wall-clock time.
Status¶
status |
Meaning |
|---|---|
pending |
No window scored yet. |
enrolling |
Neither template complete yet: fewer than 4 typing windows (about 200 keys) and fewer than 4 mouse windows (about 130 strokes). |
monitoring |
Scoring; no alarm. risk_level: normal. |
alarm |
A threshold was crossed: risk_level high (strict) or elevated (sensitive only). |
unsupported_device |
Not a desktop session (touch or mobile). Not scored yet. |
unsupported_key_mode |
Key in class mode and no mouse scoring available. With mouse scoring on, class-mode sessions are scored on mouse only. |
Limits
Desktop sessions only. Mouse scoring uses the session's own template for now; each user's mouse windows are
stored so a history template can follow. A new user's first sessions use the session typing template only,
which catches fewer takeovers (roughly half as many at the strict threshold) until history builds up. A session
needs about 200 keys of typing before keystroke monitoring starts. Validated on public datasets; your users'
false-alarm rate can differ, so check alarms on known-genuine sessions before acting on alarms.