A native Linux libfprint driver for the EgisTec (LighTuning) EH576 fingerprint sensor — a match-on-host image sensor that ships in a number of laptops but has no vendor Linux driver.
With this driver the sensor works for real: fprintd enrollment, and fingerprint
login / sudo / screen-unlock through PAM, on stock GNOME/KDE.
Heads-up before you rely on this: parts of this driver are reverse-engineered from Egis' proprietary Windows driver, including the fingerprint matcher. Please read
PROVENANCE.md— it explains exactly what is original work, what is reverse-engineered, the licensing, and the (honest) legal grey area.
| Enroll / verify / identify | ✅ works via fprintd |
PAM login, sudo, unlock |
✅ works (sufficient, password fallback intact) |
| Cross-reboot matching | ✅ (per-boot flat-field) |
| Accuracy (shipped matcher) | It admits false accepts on every dataset but one. Under the driver's own rule (two consecutive frames over the threshold): the reference unit's session of 2026-09-13 — the one every matcher constant was tuned on — 0 of 60 false rejects, 0 of 480 false accepts; the same unit, person and fingers five days later: 13 of 60 false rejects, 18 of 480 false accepts (right thumb taken for right index, where the vendor matcher on the same frames accepts none); enrolled on one of those sessions and tried with the other, 9 and 15 of 240 false accepts; sam-dant's unit 2 of 60 and 2 of 480. On irvingpop's unit 58 of 480 under the kit's looser press-maximum rule; on the second unit, under its owner's protocol, 6.9 % false rejects at 1.1 % false accepts. The cause is what the matcher measures (ridge flow and period, not identity), not any constant measured so far — see 2026-09-26: what prevents a universal 0/0. Until that date this row led with the 2026-09-13 0 / 0 as the reference unit's result. "Impostor" is always the same person's other fingers: no false-accept figure in this project comes from a second person, and physical artefacts were never tried. |
| Validated on | five laptop models (AMD + Intel; Fedora, Arch and Ubuntu). See "Tested platforms" below for what each report covered and what it left untested. |
The EH576 is a small 70×57 px capacitive/optical image sensor speaking a
simple EGIS/SIGE command protocol in the clear: the host pulls raw frames and
does feature extraction + matching in software.
This driver implements the whole chain natively: sensor bring-up → frame capture →
per-boot flat-field + preprocessing → Egis' extractor/matcher → libfprint
enroll/verify. Design detail lives in the source comments and
PROVENANCE.md.
The sensor also has a TLS-PSK session mode, reachable with class request 0x21/9 wValue=0. This driver deliberately does not use it, and neither does the
vendor's own Windows driver for this device — verified against a USB capture of a
working Windows session, which contains no vendor or class control transfers at
all — only plaintext EGIS/SIGE bulk traffic. (The mode switch is a class
request, so ruling out vendor requests alone would not have settled it.)
Entering that mode is a one-way door: the sensor stays in it across USB
autosuspend, a USB port reset and a reboot, and only a ForceResetDevice
(wValue=0x00ff, which re-enumerates the device) brings it back. On a dual-boot
machine that left Windows unable to start the sensor at all — Device Manager
Code 10 — until the board was fully powered down. Driving the same command
sequence in the clear also measures better: finger/no-finger frame variance
1069/140 = 7.6×, against 890/160 = 5.6× through the TLS path on the same unit.
The driver is engineered to be device-independent, and every per-device dependency is handled at runtime rather than baked in:
- Init / protocol — per-model constants, identical for every EH576.
- Exposure — a per-device closed-loop calibration (binary search over register
0x0fto a fixed no-finger exposure target) is measured once per fprintd process and re-applied after every sensor bring-up (the vendor init sequence resets the register, and the sensor does not retain it), so each unit self-adjusts. - Fixed-pattern noise — removed by a per-boot host-side flat-field captured fresh on every unit.
- On-chip background/vdm — neutralized to a model-level constant so no single unit's calibration data is embedded.
Update — it works on other units. Originally all of this was validated on
the author's single unit; for functional operation that caveat is retired.
Independent testers have run the full stack on other EH576 units (different
laptops, CPU vendors and USB host controllers; same sensor revision,
bcdDevice 15.72): the first with a 17-point PASS
(#2), the rest
in the table above. More reports are still very welcome — success or failure.
Accuracy is a separate axis, and there the caveat stands. Three accuracy
runs now exist — the author's plus sam-dant's and irvingpop's in
#5 — and they
disagree sharply: with the vendor matcher, 0, 2 and 37 of 60 genuine presses
were rejected at the shipped threshold (on the worst run 33 of those 37 scored
exactly 0). Those three vendor figures are from before kit_version 4 and
flatter the vendor matcher: it adapts while it verifies, and the kit loaded
each gallery once per fold where the driver reloads it for every action.
Re-measured with kit v4, the reference unit's 0 of 60 holds; the other two
have not been re-run and are lower bounds
(kit v4,
re-measurement).
Each run is one person on one unit in one session, so person and
unit change together and nothing in the data says which of the two the spread
comes from. Exposure calibration and template size both order the runs the
wrong way; the one measurement that does track the failure is how often the
matcher found an enrolment frame already covered by the template — 19, 10 and
none across the three runs — which is suggestive and nothing more. What this means practically: the driver working on your machine does not
imply the accuracy figures from another machine, so if you are going to rely
on fingerprint login, try it before you do — and please
run the kit and post the result. Detail and all
three tables: docs/matcher-comparison.md.
Note that the worst of the three runs is on a unit whose functional report in
the table below is a pass, though a qualified one (GDM login not tested there);
the kit measures an offline protocol with thinner enrolment than the driver's,
not the driver's own login success rate.
| Laptop | Platform | Sensor bcdDevice |
Distro | Result | Report |
|---|---|---|---|---|---|
| Lenovo Yoga 7 14ARB7 | AMD Ryzen 7 6800U | 15.72 | Fedora 44 | ✅ full stack (enroll, verify, GDM, sudo, unlock, suspend/resume) | author |
| Lenovo Yoga 7 15ITL5 | Intel Core i5-1135G7 | 15.72 | Fedora 44 | ✅ full stack, 17-point PASS incl. 3× suspend/resume | #2 |
| Lenovo IdeaPad Flex 5 16IRU8 | Intel Core i7-1355U | 15.72 | Arch Linux (Omarchy 4.0.2) | ✅ full stack (enroll, verify, sudo/PAM, polkit, lock screen, reboot, 3× suspend/resume); the long-lock stale-claim failure from the original report is gone on v0.4.5, re-tested over a 10 h 46 min locked screen and 1265 verification timeouts with zero refused claims | #3 |
| Lenovo Yoga 6 13ALC6 | AMD Ryzen 7 5700U | 15.72 | Ubuntu 26.04 | ✅ enroll, verify, sudo/PAM, polkit, unlock, suspend/resume (GDM login not tested) | #4 |
| Lenovo IdeaPad Flex 5 14ITL05 | Intel (11th gen) | 15.72 | Ubuntu 26.04.1 (earlier Zorin OS 18.1) | ✅ v0.5.1 Ubuntu TOD package, the vendor build: enroll, verify (the first verify right after enrolling did not match; a later run of five all matched), sudo, screen-lock unlock, unlock after suspend/resume from the menu and via the lid (2026-09-26), with the sleep hook and activation gate installed; login after a reboot and long-term use on v0.5.1 not yet tested. The gate is the reporter's design and ships since v0.4.5 (on v0.4.3, with the hook alone, unlock after suspend was not offered). Before that, v0.4.5 in daily use: stable on Zorin OS 18.1 (2026-09-20), no failure seen on Ubuntu 26.04.1 (2026-09-26) |
#3 |
Got a different machine with an EH576? Please open an issue with your results.
Upgrading to v0.5.x from any earlier release? Re-enrol every finger. v0.5.0 changed how a print is stored (a versioned record instead of the bare template bytes every release up to v0.4.5 wrote) and changed the default matcher with it, so the driver refuses prints enrolled under an earlier release instead of scoring them, and verifying against one fails. This holds for every package, the Ubuntu TOD module included, although that one still builds the vendor matcher. Run
fprintd-delete $USER, then enrol again.
Upgrading from a release before v0.4.0? Those builds spoke a TLS-PSK transport and left the sensor in a mode it keeps until something resets it — so the first fingerprint attempt after the upgrade can fail once: the driver finds the sensor in that mode, issues
ForceResetDeviceand fails that open; the sensor stays on the bus until the next transfer to it, then drops off and re-enumerates (~0.4 s), after which attempts work normally. Details indocs/sensor-tuning.md§6. That step to v0.4.x needed no re-enrollment: existing prints kept matching, because the exposure target and register are unchanged. Going on to v0.5.x does need one, see the note above.The same change fixes dual-boot: earlier releases left the sensor in a state where Windows' own driver failed to start it (Device Manager Code 10) until the machine was fully powered down. v0.4.0 never enters that mode.
Upgrading from any release before v0.4.3 (v0.1.0 – v0.4.2)? Those releases could hang: a fingerprint verify that never answers, and from then on every attempt refused with
"Device was already claimed"untilfprintdis restarted (the lock screen falls back to the password). The cause was a lost wakeup between the driver's capture thread and the fprintd main loop inside gusb's synchronous transfer helpers, narrow enough to survive months of use before it was caught on hardware. v0.4.3 fixes it — the capture thread no longer depends on the main loop at all (write-up indocs/worker-thread.md). That fix needed no re-enrollment, since capture and matching were unchanged; v0.5.x does, see the first note above.
How the driver reaches fprintd depends on the distribution. On Fedora, Arch and
Debian it is compiled into libfprint-2.so, so installing it means installing a
libfprint that includes it, replacing the distro's; the stock fprintd then loads it
through the unchanged ABI (no fprintd rebuild). On Ubuntu nothing is replaced:
every Ubuntu series ships libfprint-2-tod1, and the package in
packaging/ubuntu-tod/ builds the driver as a loadable TOD
module that the distro's own libfprint picks up
(#9, contributed by
irvingpop).
Fedora (COPR):
sudo dnf copr enable philippdev5396/libfprint-egis0576
sudo dnf upgrade 'libfprint*'
sudo systemctl restart fprintdThe COPR owner name is lower-case; dnf copr enable PHILIPPDEV5396/… fails
with a 404. As of October 2026 COPR serves v0.4.5, not v0.5.1: its newest build is
libfprint-1.94.100-99.fcNN.egis10 from 2026-09-14 (the vendor matcher and
the v0.4 template format); v0.5.0 and v0.5.1 have not been built there.
v0.5.1's Fedora package is on the
release page,
built for Fedora 43 only, and because the spec's Release: was not bumped
after v0.4.5 it carries the same name, version and release as COPR's Fedora
43 build, so dnf install of it over that build does nothing. On Fedora 43:
curl -LO https://github.com/PHILIPPDEV5396/libfprint-egis0576/releases/download/v0.5.1/libfprint-1.94.100-99.fc43.egis10.x86_64.rpm
sudo dnf reinstall ./libfprint-1.94.100-99.fc43.egis10.x86_64.rpm # over the COPR build
# sudo dnf install ./libfprint-1.94.100-99.fc43.egis10.x86_64.rpm # over Fedora's own libfprint
sudo systemctl restart fprintdthen re-enrol (first note above). For Fedora 44 and later there is no v0.5.1 package; build from source below.
Debian 13 (trixie), Ubuntu 24.04 (noble) and 26.04 (resolute): download
the .deb for your system from the
release page
(check it against SHA256SUMS there) and install it with apt:
# Ubuntu: a TOD module next to the stock libfprint, nothing replaced
sudo apt install ./ubuntu-resolute-libfprint-2-tod1-egis0576_0.5.1.26.04_amd64.deb # 26.04
sudo apt install ./ubuntu-noble-libfprint-2-tod1-egis0576_0.5.1.24.04_amd64.deb # 24.04
# Debian trixie: replaces libfprint-2-2 (never install these on Ubuntu)
sudo apt install ./debian-trixie-libfprint-2-2_1.94.100-99egis0.5.1_amd64.deb \
./debian-trixie-libfprint-2-dev_1.94.100-99egis0.5.1_amd64.deb \
./debian-trixie-gir1.2-fprint-2.0_1.94.100-99egis0.5.1_amd64.deb
sudo systemctl restart fprintdThey are files, not a repository, so apt upgrade will not update them;
install the next release's .deb the same way. The Ubuntu and the Debian
packages both install the suspend/resume hook and its gate. The Ubuntu module
builds the vendor matcher, the Debian package the clean-room default. A real
finger has been reported on the 26.04 module only; on 24.04 only the sensor
bring-up was tried on hardware
(packaging/).
Arch (PKGBUILD, not on AUR yet):
git clone https://github.com/PHILIPPDEV5396/libfprint-egis0576.git
cd libfprint-egis0576/packaging/aur
makepkg -si
sudo systemctl restart fprintdUnlike the Fedora package, the PKGBUILD does not install the suspend/resume
sleep hook and its fprintd gate — install them by hand, see
integration/.
Packaging sources and how they're published live in packaging/.
The Fedora, Arch and Debian packages replace the stock libfprint (same soname);
your other fingerprint hardware keeps working because the build includes the full
upstream default driver set. Since libfprint 1.94.100 that set includes the
egis_etu905 driver (1c7a:05ae / 1c7a:9201) Fedora used to carry as a
downstream patch, so the Fedora build no longer drops it.
No package for your distribution? Build from source below.
As noted under Install, the driver is compiled into libfprint-2.so,
so this builds a full libfprint that includes it; fprintd is not rebuilt.
You need the usual libfprint build stack, which includes OpenSSL ≥ 3.0 —
upstream's own uru4000 driver requires it (meson.build: 'uru4000' : [ 'openssl' ]).
The egis0576 driver itself links no crypto at all; it speaks the sensor's plaintext
protocol over gusb.
- Fedora:
sudo dnf install git meson ninja-build gcc pkgconf-pkg-config glib2-devel libgusb-devel openssl-devel gobject-introspection-devel nss-devel systemd-devel libgudev-devel pixman-devel cairo-gobject-devel gtk-doc - Arch:
sudo pacman -S --needed git meson ninja gcc pkgconf glib2 libgusb openssl gobject-introspection nss systemd-libs libgudev pixman cairo gtk-doc - Debian/Ubuntu:
sudo apt install git meson ninja-build build-essential pkg-config libglib2.0-dev libgusb-dev libssl-dev libgirepository1.0-dev libnss3-dev libsystemd-dev systemd-dev libgudev-1.0-dev libpixman-1-dev libcairo2-dev gtk-doc-tools - openSUSE:
sudo zypper install git meson ninja gcc pkgconf glib2-devel libgusb-devel libopenssl-devel gobject-introspection-devel mozilla-nss-devel systemd-devel libgudev-1_0-devel pixman-devel cairo-devel gtk-doc
./install.shThis clones libfprint 1.94.100, applies the driver, builds the full default
driver set (so any other fingerprint hardware you have keeps working), and installs
to /usr/local (override with PREFIX=/usr ./install.sh).
Make sure fprintd loads this libfprint, not the distro's:
ldconfig -p | grep libfprint-2 # should point at your PREFIX
sudo systemctl restart fprintdIf it still points at the distro copy, add PREFIX/lib (or lib64) to
/etc/ld.so.conf.d/, run sudo ldconfig, and restart fprintd. (Installing with
PREFIX=/usr replaces the distro library outright — simplest, but a distro update
can overwrite it, so you'd re-run install.sh after libfprint updates.)
The matcher, and the two other flavours. Since v0.5.0 the driver ships
the clean-room matcher by default: own LGPL code — an
orientation-selective Gabor front-end, a per-pixel coherence mask, a rotation
search and a ridge-period consistency check
(driver/egis0576/gabor/) behind Thaddeus
Stepanovich's adapter interface — with no vendor code in the matching path at
all. It is the flavour prepared for libfprint (that submission is on hold, see
below), and the one the numbers below are measured on: on one session of the
reference unit (2026-09-13, the session all its constants were tuned on) its
genuine and impostor populations do not overlap (0 % / 0 % at its
threshold, cross-fold checked — "impostor" means the same person's other
fingers; no false-accept figure in this project comes from a second person
yet); on no other dataset do they. On a second unit, Thaddeus Stepanovich's,
it takes his matcher from 51.7 % to 6.9 % false rejects held-out, without
reaching a clean gap there.
That clean gap is a property of one session — not of the reference unit,
and not of the matcher (2026-09-26). From 2026-09-22 until then this
paragraph called it a property of the reference unit; before that the README
presented it as the matcher's result. The same unit, person and fingers five
days later give 18 false accepts of 480 under the driver's own rule (two
consecutive frames over the threshold, the second corroborated on the
un-flat-fielded bytes), at a same-person impostor score of 0.896 against the
accept point of 0.78, where the vendor matcher on the identical frames
accepts none. Every foreign unit reported so far also scores a same-person
impostor above 0.78: 0.815 on Stepanovich's, 0.845 on sam-dant's (2 of 480
under the driver's rule), 0.903 on irvingpop's (12.1 % false accepts under
the kit's press-maximum rule, 60 % false rejects — a second failure mode,
not yet explained, in which the vendor matcher loses genuine presses too).
The reason, shown on the reference unit's own frames, is what the matcher
measures: whether two patches agree in ridge flow and ridge period, which two
different fingers can do when they land at the same angle — not whose finger
it is. None of the constants measured fixes that: threshold, overlap floor
and search width each only trade false rejects for false accepts (the
front-end's filter constants were not swept; the mechanism argues against
them, which is an argument, not a measurement). The project's target is now
an own matcher that is at least as good as the vendor one — no false accept
under the driver's rule on both reference-unit sessions, across them, on
sam-dant's and on tsteppy's data; no more false rejects than vendor on the
same data; tuned on one dataset and certified on the others, never on the
same session — and the upstream submission is on hold until one exists.
Measurements and root cause:
docs/matcher-comparison.md.
Treat the 0 % / 0 % above as what one favourable session can tell you, and
no more.
Two other flavours build from the same tree, for comparison rather than use:
EGIS0576_MESON_ARGS="-Degis0576_matcher=vendor" ./install.sh # Egis' own matcher
EGIS0576_MESON_ARGS="-Degis0576_matcher=cleanroom" ./install.sh # his original front-endvendor is Egis' extractor and matcher, machine-translated from the Windows
driver — what this driver used until v0.4.x, and what keeps it out of upstream
libfprint. On the three accuracy-tested units it rejected 0 %, 3.33 % and
61.7 % of genuine presses and scored 0 on all 480 impostor comparisons of each
run. Those false-reject rates predate kit_version 4 and are too favourable
to vendor (see "Accuracy is a separate axis" above): re-measured, the
reference unit's 0 % holds, the other two have not been re-run and are
lower bounds, and on the
reference unit's second session (2026-09-18) kit v4 gives 4 of 60 under the
driver's rule where the earlier kit gave 2. cleanroom is his original
front-end behind the same adapter: at its
published threshold it rejects 35 %, 73.3 % and 93.3 % of genuine presses on
those same captures (5 false accepts of 480 on one unit). It is not the only
flavour with measured false accepts: the default produced 58 of 480 on the
third unit under the same press-maximum rule, and 18 of 480 on the reference
unit's own second session under the driver's rule. vendor is the only one that
has never accepted an impostor press on any unit measured so far — at the cost
of the 61.7 % (or more) above. All three are measured on identical captures in
docs/matcher-comparison.md; reproduce on your
own unit with tools/accuracy/.
Templates do not survive a flavour switch (nor the v0.4 → v0.5 upgrade):
re-enrol after either. Fedora: the spec in
packaging/fedora/ builds the default; rpmbuild --with vendor builds the vendor flavour.
fprintd-enroll # or: Settings → Users → Fingerprint Login
fprintd-verify # test itEnroll with coverage: press the same finger at deliberately varied positions (centre, up/down, left/right, slight rotations) across the enrollment stages. On a sensor this small, a template built from one position matches poorly.
These steps assume
libfprint+fprintdare installed and your fingerprint is already enrolled.
Enabling fingerprint login means adding one line to the PAM auth stack:
auth sufficient pam_fprintd.so
- It must sit above
pam_unix.so(the password module). PAM evaluates auth modules top-to-bottom. sufficient= if the fingerprint matches, auth succeeds immediately; if it fails, times out, or nothing is enrolled, PAM falls through to the password. This keeps your password as a working fallback.- Do not use
requiredforpam_fprintdunless you want fingerprint-AND- password (2-factor) and accept that a missing/broken reader can lock you out. For normal "fingerprint OR password", usesufficientand keep thepam_unixline. fprintdis D-Bus/systemd-activated on all distros — it starts on demand, sosystemctl enable --now fprintdis normally unnecessary.- Keyrings still need your password. GNOME Keyring / KWallet are unlocked by the login password, so a fingerprint-only login still prompts for the keyring once.
sudo authselect enable-feature with-fingerprint(fprintd-pam provides the module.) Wires pam_fprintd.so into the system-auth /
password-auth stacks for login, sudo, and the display manager.
- Packages:
fprintd(provides the daemon andpam_fprintd.so) +libfprint. - Arch has no PAM helper — edit by hand. Add to the top of the auth section of
/etc/pam.d/system-local-login:Forauth sufficient pam_fprintd.so auth include system-loginsudo, add the sameauth sufficient pam_fprintd.soat the top of the auth section of/etc/pam.d/sudo. - GDM and KDE/Plasma have built-in fingerprint handling (don't edit
kde-fingerprint). - Keep an open root shell while editing PAM so a mistake can't lock you out.
- Ref: ArchWiki – fprint
sudo apt install fprintd libpam-fprintd
sudo pam-auth-update --enable fprintd # or run `sudo pam-auth-update` and tick the boxInstalling libpam-fprintd registers the profile but does not enable it —
the command above inserts auth sufficient pam_fprintd.so into
/etc/pam.d/common-auth (included by login, sudo, the DM). Add --force if it
refuses due to local PAM edits; disable later with --remove fprintd.
Ref: Debian Wiki – fingerprint authentication
sudo zypper install fprintd fprintd-pam
sudo pam-config -a --fprintd # disable later: pam-config -d --fprintdAdds auth sufficient pam_fprintd.so to the pam-config-managed common auth stack.
Prefer pam-config over hand-editing (it regenerates those files).
Ref: openSUSE SDB – Using fingerprint authentication
The sensor is brought up at device-open. Across a system suspend — especially
s2idle, where the USB device stays powered and is not re-enumerated — it comes
back with its capture pipeline unusable. On resume the next fingerprint attempt
(typically the lock screen after wake) can then block on it and, on some systems
(observed on AMD Rembrandt laptops that only offer mem_sleep=s2idle), hang the
unlock screen hard.
Two separate things are at play, and it took a long investigation to tell them
apart (written up in full in docs/suspend-resume.md):
- The hard freeze is prevented in the driver: every bulk read is bounded by a timeout, so talking to an unresponsive sensor fails fast to a clean error instead of spinning forever.
- "No fingerprint offered after resume" (the lock screen drops to password) is
not this driver's bug at all. It is a well-known, still-unfixed
gnome-shell/fprintd issue:
fprintdkeeps a stale device claim across suspend ("Device was already claimed"), so the post-resume claim is refused — libfprint is never even reached. See gnome-shell #7791 and Ubuntu #2067135 (it hits ThinkPads, Framework laptops, etc., on unrelated readers).
The community-standard fix — restart fprintd around suspend — is what the
helpers in integration/ do (a systemd-sleep hook that stops
fprintd before sleep and re-enumerates the sensor on resume, plus a gate that
holds fprintd's start until the sensor is back). install.sh installs them and
you should keep them installed; they are the correct fix for the upstream bug, not a workaround for a
driver shortcoming. Details and the full investigation:
integration/README.md and
docs/suspend-resume.md.
-
The wire is plaintext, by design. Traffic to the sensor is the vendor's own
EGIS/SIGEprotocol, exactly as the shipped Windows driver sends it, and the driver links no crypto. The sensor does have a TLS-PSK session mode; using it would not have bought a security property here — the key is a fixed per-model constant, the channel would terminate in the same userspace process that then handles the decrypted image anyway, and matching is host-side, so the template is on disk regardless. On Linux the meaningful boundary is match-on-chip, which this sensor does not offer. The variance figures quoted above, and what else was measured and found to be at its limit, are indocs/sensor-tuning.md. -
Nothing blocks the fprintd main loop. The transport is a set of
FpiSsmmachines overFpiUsbTransfer— the in-tree libfprint model — and only the matcher runs on another thread (aGTask, assecugendoes).open()completes asynchronously too. Cancellation is honoured between transfer sequences and never handed to gusb: aborting an in-flight URB wedges this sensor until board power is cut (measured). The worker thread of v0.4.x and the lost-wakeup hang it once had are history, kept indocs/worker-thread.md. -
Press firmly, flat, centred, and hold ~2 s. Verification scores every frame while the finger is down and takes the best; light or brief taps on a 70×57 sensor can score zero. How much this matters clearly differs between people and units: in one of the three accuracy runs 33 of 60 genuine presses scored exactly 0 (
docs/matcher-comparison.md). -
The match threshold is strict and not user-tunable by design — the priority is rejecting impostors (including adjacent same-hand fingers), not maximum convenience. Lowering it would not have rescued that run either: 33 of its 37 failing presses scored 0, which no threshold recovers.
-
Templates are stored by fprintd under
/var/lib/fprint/, protected by filesystem permissions (as with every libfprint driver). -
Re-enroll after anything that changes capture conditions materially (the exposure calibration converges to the same fixed target every boot, so normal reboots are fine).
-
The first verify (or two) after upgrading fails, then it works. Expected, once: the driver found the sensor in the session mode an older release left behind and issued a reset; the sensor re-enumerates on the next access (
docs/sensor-tuning.md§6). See the note under Install. -
Every verify fails right after upgrading to v0.5.x: prints enrolled under v0.4.x or earlier are refused, not scored. Re-enrol; see the first note under Install.
-
fprintdstill uses the old driver / device not found: confirmldconfig -p | grep libfprint-2points at yourPREFIX, thensudo systemctl restart fprintd. -
Enrollment "fails to capture": press more firmly and hold; ensure no finger is on the sensor during the first ~2 s (baseline + exposure calibration run then).
-
Verify always fails after it worked: re-enroll with coverage (see "Enrolling").
-
Debug logging:
sudo G_MESSAGES_DEBUG=all fprintd(stop the service first) orjournalctl -u fprintd -f.
Most wanted, in order:
- A clean-room fingerprint matcher good enough to distinguish adjacent same-hand
fingers at 70×57 — this is what blocks clean libfprint upstreaming. The
shipped Gabor matcher does not (it compares ridge flow and period, not
identity); the bar a replacement has to clear, and why, is in
docs/matcher-comparison.md. See the "Why the reverse-engineered matcher (and not a clean-room one)?" inPROVENANCE.md. - Test reports from other physical EH576 units (see "Universality").
- Accuracy reports from your own fingers — especially if the result is
poor; the three runs collected so far disagree sharply and more points are
the only way to understand that. Run the
tools/accuracy/kit (about 15 minutes; only theresults.jsonsummary is shared, never frames) and open an issue titledaccuracy: <laptop model>as its README asks — collected under #5. - Packaging. The release CI and the Debian/Ubuntu
.debpackages landed through #6 (closed 2026-09-21): the release workflow builds, and v0.5.1 carries, a Fedora 43 RPM, an Arch package, Debian trixie.debs and Ubuntu noble/resolute TOD.debs. Still wanted: thePKGBUILDpublished on the AUR (it is not there); an apt repository or PPA for the.debs (today they are release-page files thatapt upgradenever updates, and the PPA named inpackaging/does not exist yet); COPR rebuilt on every release tag, since it is the only source of Fedora 44 and later packages and is still at v0.4.5 (the release workflow never pushes there); a way to rebuild automatically when a distro updates libfprint; recipes for other distributions.
Original driver code: LGPL-2.1-or-later (see LICENSE), matching
libfprint. Reverse-engineered portions and the full legal picture:
PROVENANCE.md. The original Egis Windows driver and its
decompilation are not included and must not be committed here.
Reverse-engineered and ported to libfprint by the repository author, with the goal
of making already-owned hardware usable on Linux. The optional clean-room
correlation matcher (driver/egis0576/tsteppy/, LGPL-2.1-or-later) is the work of
Thaddeus Stepanovich. The suspend/resume activation gate in
integration/ implements a design diagnosed, built and tested by
sam-dant. The Debian and
Ubuntu packaging, the Ubuntu TOD module and the release CI's version checks were
contributed by irvingpop, who maintains that part of
the packaging stack. Prior art
this project started from is listed in
PROVENANCE.md §3; the people who tested the
driver on their hardware are named in the "Tested platforms" table above. Built on
the excellent libfprint and
fprintd.