Jorijn van der Graaf
ba882ee47b
Jorijn asked for the control that settles it -- a fresh template, tested with learning off -- and ran it twice. fresh template, 0 folds 30/30, two consecutive blocks of fifteen same lineage, 40 folds 12/15 same lineage, 185 folds total failure, 108 consecutive rejections Every one of those measured with learning switched off during the measurement itself, so nothing moved underneath the numbers, and the fresh-template result is replicated back to back. Three points, monotonic in fold count. The mechanism has been visible since the 185-fold collapse: the frames one press contributes are near-duplicates of a single image from one finger position, so folding them spends the template's ninety-six slots on that position and evicts the diversity a twenty-sample enrolment put there. Stock's updates are spread across many separate presses hours apart, which is where diversity actually comes from. And there is nothing on the other side of the scale. A plain enrolment measures thirty out of thirty, so learning has no headroom to improve anything, and it has never once been observed to raise a rate under conditions worth defending -- the run that once looked like proof was confounded by a freshly wiped sensor and a user learning the technique, both of which Jorijn identified himself while the numbers were still climbing. The code stays behind --learn=1. The finding is about this trustlet's algorithm, not about the idea.
176 lines
8.9 KiB
Markdown
176 lines
8.9 KiB
Markdown
# fingerprintd
|
|
|
|
Fingerprint daemon for the Fairphone 6 (`milos`, SM7635) on mainline Linux.
|
|
|
|
## Why a daemon
|
|
|
|
The sensor is a FocalTech FT9391 on a TrustZone-owned SPI bus. `spi@a88000` is
|
|
`disabled` in both the mainline and the stock Android device tree, and the pads
|
|
are XPU-protected — touching them from the normal world is an instant SError
|
|
reboot. Every pixel the sensor produces stays inside the TEE: capture,
|
|
preprocessing, the classifier, enrolment and matching all run in the `focal64`
|
|
trustlet, which reports a matched finger id and nothing else. A libfprint-style
|
|
driver cannot exist on this device.
|
|
|
|
So the normal world's job is narrower than usual, and none of it is per-request
|
|
work:
|
|
|
|
* **Power the sensor.** Rail on gpio29, reset on gpio74, interrupt on gpio75 —
|
|
the same division of labour the downstream driver uses. One sensor reset buys
|
|
exactly one trustlet init, so whatever powers the sensor must also hold the
|
|
session open.
|
|
* **Be QTEE's filesystem.** QTEE cannot reach storage. When the trustlet saves
|
|
or loads a template it calls back into the normal world through the `gpfile`
|
|
(0x7000) and RPMB (0x2000) listeners, and expects them served. QTEE does the
|
|
crypto and the anti-rollback; this side moves opaque bytes and performs the
|
|
authenticated RPMB transactions against the UFS device.
|
|
* **Speak a biometrics API.** The daemon owns `net.reactivated.Fprint`, so
|
|
`pam_fprintd`, the Plasma fingerprint KCM and `fprintd-enroll(1)` work
|
|
against it unmodified.
|
|
|
|
A listener registration is held for as long as the process lives and QTEE's
|
|
listener table is global to the boot, so this has to be one long-lived process
|
|
rather than a tool spawned per request.
|
|
|
|
## Layout
|
|
|
|
```
|
|
interfaces/ Fingerprintd{,-Sfs}.cppm the core: pure C++ modules
|
|
implementations/main.cpp the daemon shell
|
|
tests/ one suite per core module
|
|
```
|
|
|
|
`fingerprintd-core` is a static library with no GLib, no libqcomtee and no
|
|
system headers. Everything in it is a wire format or a state machine that was
|
|
recovered by reverse-engineering, so all of it is pinned by tests that run on a
|
|
dev box with no phone, no TEE and no sensor. The daemon shell holds everything
|
|
that touches hardware.
|
|
|
|
## Build
|
|
|
|
```sh
|
|
crafter-build # bin/fingerprintd-<target>-<march>/fingerprintd
|
|
crafter-build test # the unit suites
|
|
```
|
|
|
|
Cross-compiling for the phone:
|
|
|
|
```sh
|
|
packaging/make-sysroot.sh # once; no root, no qemu, no device
|
|
crafter-build -- --target=aarch64-alpine-linux-musl \
|
|
--sysroot=~/.cache/fingerprintd/sysroot-aarch64-alpine \
|
|
--march=armv8-a --mtune=generic
|
|
crafter-build test --target=aarch64-alpine-linux-musl --sysroot=... \
|
|
--march=armv8-a --mtune=generic # runs the suites under qemu-aarch64
|
|
```
|
|
|
|
The result links dynamically against the phone's own musl and libc++
|
|
(`libc++`, `libc++abi`, `libunwind`, `libgcc_s`, all already present on pmOS).
|
|
The research harness this replaces had to be built `-static`, but only because
|
|
it was built with the host's glibc toolchain — that constraint does not apply
|
|
to a real Alpine sysroot.
|
|
|
|
Verified on the device: all five suites pass cross-built and run **on the phone
|
|
itself**, not only under emulation.
|
|
|
|
## Status
|
|
|
|
**It works end to end through fprintd's own clients.** On the Fairphone 6, as a
|
|
systemd unit owning `net.reactivated.Fprint`:
|
|
|
|
```
|
|
fprintd-enroll -f right-index-finger user ten stages, enroll-completed
|
|
fprintd-list user - #0: right-index-finger
|
|
fprintd-verify user (wrong finger) verify-no-match, on the first press
|
|
fprintd-verify user (enrolled finger) verify-match, on the first press
|
|
```
|
|
|
|
Nothing in fprintd was modified; the daemon speaks its interface.
|
|
|
|
| module | what it holds |
|
|
|---|---|
|
|
| `:Sfs` | the gpfile frame — the read/write offset split, the `O_TRUNC` guard, root mapping, path-traversal rejection |
|
|
| `:Rpmb` | request/reply framing, the bytes-transferred out-parameter, JEDEC result codes, chunking, the one-time-programmable key guard |
|
|
| `:Ta` | command surface, the 740-byte event context, capture flags, save masks, enrol/auth payloads, the error table, the verdict rule |
|
|
| `:Engine` | baseline calibration, touch edges, enrolment progress, the accounting |
|
|
| `:Store` | the finger name map |
|
|
| `:Tee` | QTEE service/op numbers, the listener table, the 13-byte CBOR credentials blob |
|
|
| `:Sensor` | the pins, the timings, and the XPU guard |
|
|
|
|
Every constant that was recovered by reverse-engineering carries where it came
|
|
from, and the tests are written to fail if it is undone rather than to restate
|
|
it. Several replay real captures off the phone.
|
|
|
|
Two policy decisions live in the daemon and both were forced by measurement.
|
|
Verification is judged **per press**: any matching frame wins, only rejections
|
|
is no-match — both correct presses in the acceptance run had rejected frames
|
|
before the one that matched, so a first-frame rule would have failed them. And
|
|
the shipped config sets `max_authentication_rescan_times` to 0, because at the
|
|
stock budget a wrong finger never yields a terminal frame and a PAM client
|
|
waits forever for the `verify-no-match` it needs.
|
|
|
|
**Template learning is implemented and OFF, because it makes matching worse on
|
|
this hardware.** Measured, with learning disabled during each measurement so
|
|
nothing moved underneath the numbers:
|
|
|
|
| template | folds | rate |
|
|
|---|---|---|
|
|
| fresh 20-sample enrolment | 0 | **30/30** |
|
|
| same lineage, later | 40 | 12/15 |
|
|
| same lineage, later still | 185 | total failure |
|
|
|
|
The frames one press contributes are near-duplicates of a single image from one
|
|
finger position. Folding them spends the template's 96 slots on that position
|
|
and evicts the diversity the enrolment put there. There is no upside to weigh
|
|
against it either: a plain enrolment already measures 30/30, and learning has
|
|
never once been observed to raise a rate under controlled conditions.
|
|
|
|
`--learn=1` enables it for experiments. What follows describes how it works.
|
|
|
|
**Template learning.** Stock rewrites the stored template on every successful
|
|
press — `0x1015 UPDATE_TEMPLATE` while the finger is still down, then a deferred
|
|
`SAVE_DATA` — and the stored body measurably grows over a template's life
|
|
(333278 bytes at enrolment to 371734 after one authentication session on the
|
|
reference device). This daemon does the same as of 0.1.0. It matters more than
|
|
any tuning knob: before it, every match rate measured against this device was
|
|
measured against a day-zero template that no stock user lives with.
|
|
|
|
The harvest sends no event, so the matcher does not re-run and a verdict cannot
|
|
be revised by it. A quick tap pays almost nothing, because the finger is gone by
|
|
the time a verdict lands; a held press contributes the frames it was held for —
|
|
which means the frames learned from are, by construction, frames from real
|
|
unlock presses. The save is deferred until after the verdict reaches the client,
|
|
because ~350 ms of RPMB traffic does not belong on an unlock path; stock defers
|
|
it the same way. `--learn=0` turns the whole thing off so the comparison can be
|
|
made on one binary, which is the only way it is single-variable.
|
|
|
|
**Reading the trustlet.** There is no `tzdbg` on mainline, but focal64 writes
|
|
its log into the response buffer, so `--ta-log` surfaces the matcher's own
|
|
verdicts (`auth success score`, `identify fail! FtVerifyByTemplate() = -2`, the
|
|
per-frame `image quality / coverage / humidity`). The log ring is ~150 lines per
|
|
session and is never reset, so the config dump alone can overflow it: the
|
|
shipped verbose config leaves the framework log off to reserve the ring for the
|
|
algorithm. Lower levels are more verbose and 6 is off.
|
|
|
|
**Known about the loop, from real use.** A verify frame is four QTEE round
|
|
trips (~200 ms idle); a `REPORT_EVENT` that runs the matcher is ~300 ms, and
|
|
the rising edge pays it twice because events 5 and 7 both reach the matcher.
|
|
The first frame of a press is the finger landing and usually rejects; matches
|
|
land at frame 3, 5, 8 — so a quick tap is often a miss. Two attempts to fix
|
|
that (drop event 5; recapture after 50 ms) went in together and produced zero
|
|
matches; both are reverted and recorded. gpio75 is a ~1 ms pulse, not a
|
|
level — an IRQ-driven idle needs GPIO edge events. Changes to this loop are
|
|
made one variable at a time, on a measured baseline.
|
|
|
|
**Not done:** packaging (`provides="fprintd=…"` so this replaces the fprintd
|
|
daemon package while `fprintd-pam` stays), the shipped storage policy, polkit
|
|
(a caller-uid rule stands in), trustlet-side template removal (deletes drop the
|
|
name only), and the kernel config change — `CONFIG_QCOMTEE`, which selects the
|
|
SHM bridge — that gates any public image. A claim held by a client that leaves
|
|
the bus is dropped.
|
|
|
|
## Runtime dependencies, not carried here
|
|
|
|
The `focal64` trustlet is proprietary and is **not** in this repo. It is
|
|
extracted from the device's own stock Android partition on first boot by the
|
|
`fp6-vendor-blobs` mechanism, the same way the audio firmware is.
|