Relay attack watch for keyless cars, on a Flipper Zero.
Leave it in the car. Tell it you are walking away. It listens for your key answering a question nobody asked.
A Flipper parked in a car can see a keyless theft happening thirty metres away, through a house wall, without hearing the part of the attack that is being relayed. That sounds wrong. Here is why it is true.
A passive-keyless car asks its key a question over a 125 kHz magnetic field from the door handle. That field reaches about a metre — it is near-field, and its strength falls off with the cube of distance. The key answers over UHF: 315, 390, 433.92, 868 or 915 MHz depending on where the car was sold. That answer is an ordinary radio transmission, built to unlock a car from across a car park, and it carries hundreds of metres.
The common, cheap theft relays only the question. One person stands at the car with a loop that picks up the 125 kHz challenge; a second stands by your front door and re-radiates it there. Your key wakes up in the hallway and answers.
Nobody relays the answer, because nobody has to. The key's UHF reply reaches the car on its own — through the wall, across the drive, exactly as it was designed to. That is what makes the attack practical with two cheap boxes.
It is also what leaves the attack audible from the driver's seat.
YOUR HOUSE YOUR CAR
┌───────────────┐ ┌──────────────────┐
│ key ((o)) │ ── UHF reply, 100s of m ───► │ car FLIPPER │
│ ▲ │ nobody relays this │ ▲ │
└────────┼──────┘ └──────────┼───────┘
│ │
125 kHz, re-radiated this is what
by the thief's box Cardea hears
▲
└──────── relayed by hand, box to box ───────────┘
So Cardea sits with the car and listens for a key that is answering when nobody asked it anything.
Guard hops every key band in the world — 315, 390, 433.92, 868.35, 915 MHz — and camps on whichever one speaks. It pulls individual bursts out of the RSSI stream and grades what it hears against three independent tests:
| Family | What it measures | Why it is evidence |
|---|---|---|
| UNSOL | Key-shaped bursts while you have declared yourself away | Nobody was there to press anything |
| RHYTHM | Mean absolute deviation of the gaps, over their mean | A relay poller is a software loop. A thumb is not. |
| CLONE | Consecutive bursts of the same length, shape and strength | The same transmission, arriving again |
Then it says one of four things: QUIET, ODD TRAFFIC, SUSPICIOUS, or RELAY LIKELY — and stops there. It never says "confirmed", because without decrypting your key's reply there is no way to prove the burst came from your key.
These are not editorial. Every one is enforced in helpers/cdr_detect.c and
pinned by a test in test/host_detect_test.c.
One burst is never evidence. Radios are noisy and your neighbour also owns a key. The unsolicited ladder scores zero at one burst and zero at none.
RELAY LIKELY needs two of the three families to agree. One family, however loud, tops out at SUSPICIOUS.
One of the two has to be RHYTHM. This rule was put there by a failing test, not by taste. Picture yourself in your own hallway pressing your own key six times: those bursts are unsolicited, because you told the app you were away, and they are perfect clones of each other, because it is one key sending one frame. Two families, sixty points, and nothing whatsoever has gone wrong. The metronome is the only one of the three that a human hand does not produce, so the top verdict is not allowed to do without it.
A metronome that has been ticking for over a minute is furniture. Weather stations, tyre sensors and doorbell repeaters are all perfectly periodic and perfectly innocent. Past sixty seconds the BEACON flag fires, RHYTHM and CLONE stop scoring, and UNSOL is halved — so a weather station settles at ODD TRAFFIC with a tag on it instead of screaming all night. The cost, stated plainly: an attacker who kept a metronome-steady poll running for over a minute would be down-weighted. No relay tool behaves that way, and the trade buys immunity to every periodic sensor on the street.
The score is capped at 92. The three ceilings sum to 90. There is no measurement here that deserves 100.
Given those weights, RELAY LIKELY can only ever be reached while armed. RHYTHM and CLONE together are 50, and the threshold is 60. That is a property of the arithmetic, and there is a randomised test asserting it over three thousand generated situations.
A held band is reported separately. A carrier sitting on a band gets its own HELD tag and is never folded into the score — partly because it may be the attacker's own link, and partly because it blinds the burst detector, which you deserve to be told rather than left staring at a reassuringly quiet screen.
- It cannot hear the 125 kHz half. Nothing on a Flipper can, at that range — the LF antenna is tuned for RFID at contact distance, not for sensing a relay coil across a driveway. Cardea watches the UHF half, which is the half that travels.
- It cannot decrypt your key's reply. Passive-entry responses are manufacturer-proprietary and rolling. Cardea matches on duration, envelope shape and band — not on content.
- It cannot prove a burst was your key, only that it was shaped like it.
- It cannot stop a theft. It tells you one is being attempted.
- A quiet screen is not proof nothing happened. It hears one half of the attack, not both, and while hopping it is listening to roughly a fifth of each band at a time. A single 40 ms frame can be missed. A train cannot.
Optional, and the single biggest improvement available. Hold your key against the Flipper and press LOCK twice; Cardea captures the burst, checks the second press agrees with the first, and averages them.
Without a learned key, "key-shaped" means anything from 4 to 150 ms on any band — honest, but broad. With one, it means one length, one envelope, one band, and every other remote in the street stops counting. Two agreeing presses are required because one reading can be a reflection off the car body and two cannot.
Put the Flipper somewhere in the car — glovebox, door pocket, under a seat — open Guard, and walk away. It arms itself after the delay you set (10 s by default) and resets its evidence window at that moment, so the walk to your front door is not counted against you.
| Key | Does |
|---|---|
| OK | AWAY / HERE — the single most important control in the app |
| Hold OK | Mute. Silences the speaker and the motor, never the LED. |
| ← → | Watch / Detail / Bands, and the alert's three pages |
| ↑ ↓ | Pin the receiver to one band, or return to hopping |
| Back | End the watch (the report keeps what it saw) |
AWAY is what makes an unexplained burst unexplained. While it reads HERE, your own key accounts for everything and UNSOL scores nothing.
Reading the watch screen. The strip is thirty seconds of history, one pixel column per quarter-second, bar height being strength over the noise floor. The dotted line is the level a burst has to clear before it counts at all — bars below it were seen and deliberately not scored. The square at top-left is filled while camped on a band and hollow while hopping.
Alerts and anything reaching SUSPICIOUS are appended to
/ext/apps_data/cardea/events.csv.
Battery. The receiver runs continuously; expect a few hours on a charge. For an overnight watch, plug it into a power bank.
Download cardea.fap from Releases
and copy it to /ext/apps/Sub-GHz/ on the SD card. It appears under
Apps → Sub-GHz → Cardea.
FAPs are firmware-API specific. This one is built against the release channel (API 87.1) and is CI-checked against the dev channel too. If it refuses to start, rebuild it against your firmware:
python3 -m pip install --upgrade ufbt && ufbt && ufbt launchufbt # build dist/cardea.fap
ufbt launch # build, install and run on a connected Flipper
make -C test # the engine's host testsThe detection engine has no Flipper headers in it, so it compiles and runs on a
laptop under the same -Werror settings the firmware uses. That is deliberate:
the reading of the signal is the whole product, and it is the one part a
screenshot cannot vouch for.
18476 checks, 0 failures
helpers/cdr_detect.{c,h} the engine: floor tracking, burst extraction,
envelopes, the three families, the scorer. Pure C.
helpers/cdr_radio.{c,h} one worker thread: hop, camp, sample, publish
helpers/cdr_store.{c,h} settings, the learned key, the CSV log
views/ splash, guard (3 pages + alert), learn, report,
primer
scenes/ start, guard, learn, report, primer, settings, about
test/ host tests for the engine
tools_gen_*.py icons, banner, screen mockups (Pillow)
The screen images in this README are rendered by tools_gen_mockups.py, which
imports the same layout constants the C uses and ports the draw code line for
line. It is not decoration — it caught four real collisions in this app before
any of them reached a device, including chip labels overflowing their own
frames and a report label running through its own value.
Cardea is receive only. It never transmits, never replays, never emulates a key, and contains no code that could. It is a passive monitor for your own vehicle and your own keys.
Passive radio monitoring is legal in most jurisdictions; acting on what you hear may not be. If it fires, the right response is to move your keys, check your car from indoors, and call the police — not to go outside and confront anybody.
Cardea is the Roman goddess of the door hinge — the one who keeps at the threshold what belongs outside it. She is the companion of Janus, which is fitting, since Janus is already in this family.
Built by at0m-b0mb. MIT licensed.