The fingerprint reader in the ThinkPad X1 Tablet Gen 3 is a Synaptics SYNA8002 on SPI. No Linux driver exists for it, and no SPI Synaptics driver exists in libfprint at all.
This is the protocol, and a userspace driver built on it that logs a real user into GDM with a thumb.
It is published as reverse engineering, not as a finished product. Read the caveats before you get excited — in particular, it currently needs a key that only exists if the machine has run Windows.
| SPI transport | ✅ solved, confirmed on hardware |
| Mutual-auth TLS handshake to the sensor | ✅ completes |
| Session setup (config blob) | ✅ accepted |
| Image capture | ✅ 144×40, ~0.3 s per frame |
| Enrolment | ✅ own tool, graded per press |
| Matching | |
| PAM / GDM login | ✅ works |
| Touch-to-wake and unlock while locked | ✅ works |
| Survives suspend/resume and cold boot | ✅ works |
| Same template accepted a day later | ✅ measured, see Results |
| Pairing a virgin sensor from Linux | ❌ not implemented — needs a Windows-derived key |
| libfprint driver | ❌ not attempted |
PROTOCOL.md the recovered protocol — transport, secure channel, capture, image format
synaspi/ the driver (Python, ~3.5k lines)
packaging/ systemd unit, D-Bus policy, install/uninstall — this is what puts it in the login path
eval_enroll.py offline scoring of a template set against probe and impostor sets
NBIS-FINDINGS.md why NBIS (mindtct/bozorth3) cannot match at this sensor area
VERIFICATION-LIMITS.md what limits verification on a worn finger — measured, mostly negative
synaspi/hold.py capture a run of frames during ONE sustained press (the tool behind it)
outreach-draft.txt drafts of the messages accompanying this release
Inside synaspi/:
| module | |
|---|---|
transport.py |
SPI framing: status/write/read, and the 0x05 bulk image path |
identity.py |
injects the recovered host key into the handshake |
bringup.py |
power cycle → handshake → identify → session setup |
configblob.py |
selects and sends the vendor config blob (blob not included) |
lineupdate.py, capture.py |
the product-0x33 capture command, and reading frames |
reconstruct.py |
raw lines → a 144×40 image |
matcher.py, template.py, quality.py |
host-side matching and enrolment templates |
session.py, daemon.py |
open once, capture many; the verify loop |
fprintd_adapter.py |
net.reactivated.Fprint on the system bus, so PAM works unmodified |
wake.py |
watches the sensor while the session is locked; press to wake and unlock |
enroll.py |
graded enrolment: scores each press and can reject it |
enhance.py |
Gabor ridge enhancement (see the caveats — it did not help) |
Talk to the sensor once, without touching anything system-wide:
sudo modprobe spidev bufsiz=65536 # the image reads are large
sudo PYTHONPATH=$PWD python3 -m synaspi.bringup
sudo PYTHONPATH=$PWD python3 -m synaspi.enroll --prefix me # then score it with eval_enroll.pyTo actually log in with it, packaging/ installs the adapter as the system's
net.reactivated.Fprint provider, which makes PAM, GDM and GNOME Settings use the
sensor with no client-side changes:
sudo SYNASPI_PV_DIR=~/src/python-validity SYNASPI_RECOVERED=~/syna-key \
./packaging/install-local.shRead packaging/README.md first — that command puts unreviewed code in your
login path, and the one thing you must have ready is a second way into the machine.
packaging/uninstall-local.sh reverts all of it.
You need two things this repository does not contain:
- Your sensor's host private key. Recovered from your own Windows install's registry,
under
SYNA8002\...\Device Data\PairingData.PROTOCOL.md§4 describes the format. PointSYNASPI_RECOVEREDat a directory holdinghost_priv.hex. - The session config blob, extracted from
synaWudfBioSpi.dllon your machine.PROTOCOL.md§5 describes the catalogue. Place it insynaspi/blobs/.
Both are per-device or vendor-owned. Neither is ours to ship.
It also depends on python-validity for the
TLS and command layer — point SYNASPI_PV_DIR at a checkout.
Numbers from this tablet, one enrolled right thumb, threshold 0.400.
Separation (natural-placement enrolment, eval_enroll.py, alpha 0.0):
| genuine floor | +0.4307 |
| impostor max | +0.3402 (21 impostor views, 5 fingers) |
| margin | +0.0905 |
| FRR / FAR at thr 0.35–0.40 | 0% / 0% |
| template self-consistency (leave-one-out) | min +0.5963, mean +0.7847 |
Cross-day, in production. The interesting question for a fixed threshold is not how a press scores two minutes after enrolment — it is how it scores tomorrow. Template enrolled one morning; the machine then spent 7 hours powered off; the next morning, at the GDM greeter on a cold boot:
06:50:16 VerifyStart(any) -> verifying right-thumb
06:50:25 press: presence 22.2 score +0.3356 (thr 0.400) reject
06:50:25 press: presence 35.1 score +0.5117 (thr 0.400) ACCEPT
06:50:25 verify -> verify-match best +0.5117, presses 2/11, 3.5 s
+0.5117 is above the genuine floor of the same-session probes, so a day of drift did
not eat the margin. Two secondary observations from the same log: presence is a usable
press-quality signal (a weak placement at 22.2 scored below threshold, a solid one at 35.1
scored well above), and a conservative threshold plus cheap retries is the right trade —
the first press failing cost 1 second, not a fallback to the password.
One data point is one data point. It is evidence that the approach is sound, not a stability claim over weeks.
It needs Windows to have been there first. Everything here reuses an existing pairing. Nothing re-pairs a virgin sensor, and we do not know whether that is possible.
The matcher is not authentication-grade, and we can prove it. The image is 144×40 —
too small for minutiae, now measured rather than assumed (see NBIS-FINDINGS.md) — so
matching is correlation-based. Measured against real impostor
fingers, the first version separated genuine from impostor by +0.0005, which is not
separation at all. Natural-placement enrolment moved that to +0.0905, with FRR 0% /
FAR 0% at threshold 0.35–0.40. That is a real separation, and it holds across days — but
it rests on small numbers and one enrolled finger.
The number that should stop you using this as your only authenticator is the false accept bound. Zero false accepts out of 21 impostor views bounds FAR at roughly 14% with 95% confidence (rule of three) — that is what 21 samples can tell you, no more. Windows Hello's small-area bar is FAR < 0.002%. Closing that gap is a question of collecting several orders of magnitude more impostor data, and quite possibly of a better matcher; it is not something this repository has done.
Two things that turned out to be false, recorded because they cost real time: more
enrolment views does not fix this — impostor scores rise faster than the genuine floor —
and Gabor ridge enhancement raises impostor scores more than genuine ones. The second was
re-tested on 2026-11-09 against a worn finger and a natural template, and still holds:
alpha=1.0 lifts the genuine median from +0.477 to +0.658 but the worst impostor from
+0.340 to +0.624, taking false rejection from 5% to 40%.
Capture strategy is exhausted, and the reason is in VERIFICATION-LIMITS.md. Ten
captures of one sustained press agree to within 0.008 and sit at 39:1 SNR, so there is no
noise worth averaging; and the sensor returns the same patch whether the finger is held
still or deliberately rolled (zero displacement, NCC 0.998 between consecutive captures),
so there is no area to mosaic either. All the variation is between presses — spread
0.346, forty times the within-press figure — and nothing observable predicts it: contact
presence correlates with score at r = -0.13. A bad landing cannot be detected early, only
scored and retried.
The unlock-on-touch path steps outside PAM. Same matcher and threshold, but PAM's attempt counting is not in the loop. There is a flag to disable it.
Timing. Bring-up is ~10.8 s, of which 8 s is the mandatory power-low. Captures are ~0.3 s once the session is open, so retries are cheap.
The obvious plan was that matching would move to NBIS, deleting our weakest component.
That plan does not work, and we have now tested it rather than assumed it. NBIS cannot
match at this sensor area: cropping clean 500 dpi reference prints down to the ~29 mm² this
sensor sees drops bozorth3 to zero on genuine pairs, and on our own captures mindtct
finds a median of 8–12 minutiae but bozorth3 scores 0 on all 120 genuine pairs. Usable
separation only appears somewhere around 130–180 mm². The limit is sensor area, not image
quality, so no amount of enhancement work changes it. Full method and numbers in
NBIS-FINDINGS.md.
So a libfprint driver for this part cannot simply hand up an image and rely on the stock
matcher — it would enrol and then never match. It needs a matcher suited to small-area
sensors (correlation-based, as here), or the device has to be driven match-on-chip.
elanspi is the only SPI precedent in libfprint and would still define the shape of the
transport. The pairing and blob questions above have to be answered first; see
outreach-draft.txt.
The secure channel is a port of python-validity
over an SPI transport. Its tls.py and command layer are what made this tractable at all.
Validity90 is the protocol work underneath that.
MIT — see LICENSE. Chosen to match python-validity, which this builds on.
Portions of this work build on python-validity, Copyright (c) Vladimir Semenov and contributors, also under the MIT License. The TLS and command layer are used from that project; the SPI transport, the SYNA8002-specific capture path, the image reconstruction, the matcher and the fprintd adapter are new here.