Product1 publisher2 min readPublished
Flock's on-camera software classifies people, bicycles and bumper stickers
Hackers pulled a camera off a roadway, recovered the encryption key stored inside it, and handed the files to 404 Media and WIRED. The joint analysis found a person detector inside an automatic license plate reader.
The Product Desk · Product desk
What happened
- Hackers ripped a Flock camera down from above a roadway, made a near-complete copy of the data stored inside it, and shared the files with 404 Media and WIRED.
- The joint analysis of the recovered data found that software running on the device explicitly detects people, along with vehicles, license plates and bicycles.
- The hackers recovered an encryption key stored on the camera itself, and it unlocked videos of thousands of vehicle detections.
- Multiple people have been arrested around the country for allegedly tampering with Flock cameras, and one police department built a fake 3D-printed camera case to bait vandals.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- decision An agency that has told a council its cameras read plates now has to either produce the on-device class list in writing or change the sentence it uses in public.
- exposure Every deployed unit is the same hardware, so physical custody of one camera is enough to reach what that camera stored, and the risk scales with the number of poles, not the strength of the network.
- precedent With the hackers intending to publish their method, a second dump becomes a planning assumption for any department running these cameras.
- constraint Because part of the storage stayed locked, an outside review still cannot establish what the camera keeps about the people it detects. The locked storage caps what any oversight report can conclude.
A computer-vision pipeline that outputs a person box is a different product from one that outputs a plate string, and a retention policy sits on top of the classifier without changing what the classifier can name. The recovered software also cut out bumper stickers and other graphics, including an American flag patch on a motorcyclist's saddlebag [8].
Volume is the part that makes an audit hard. Several weeks of logs from the single unit account for more than a million images [6], and the device can produce dozens of images of one passing vehicle [7]. Take thirty frames per pass and a million images works out to roughly 33,000 vehicle passes [14]. Read "several weeks" as four and the same total is more than 35,000 images written a day [15], from one camera on one road.
Flock has described the system as protected by on-device encryption [12]. On-device encryption protects against a remote attacker, not against someone who takes the camera down. Custody of the hardware is a separate control, and here the files went from the hackers to 404 Media and to the transparency nonprofit Distributed Denial of Secrets, which passed them to WIRED for a joint analysis [13]. Flock's answer to the findings is not in the account of the breach.
For whoever has to speak at the next public meeting, the answers are checkable in writing. First, the list of object classes the on-device model outputs, which is a build artefact. Second, images written per camera per day, which the vendor can read from the same logs the hackers read. Third, where the decryption key sits, and whether taking a camera down takes the key with it.
Answers one and three are the ones that decide whether a council-meeting sentence about plates matches the device on the pole. A policy answer to a capability question still passes the exposure to the agency, because the agency is the party named in the minutes and in the public records request. Some towns have already announced they will stop using Flock's cameras altogether [10].
What to watch
- Whether Flock publishes the on-device object class list or responds to the person-detection finding.
- Whether the hackers' published method produces a second camera dump, and in which town.
- Whether a council cancels or pauses a Flock contract and cites the classification finding in the minutes.