Vector State: License Plates for Pedestrians

An oft-repeated claim is that Flock cameras don’t detect people, and when they do it’s only basic movement detection, and totally not facial detection. At this point, would we be surprised if pretty much all of that turned out to be a lie?

Ghost in the Machine: The “Face Detection” Switch

With that out of the way, this is the list of triggers defined for Flock devices, as found in one of their web applications:

• Manual
• Common alarm
• Motion detection
• Input switch alarm
• Video loss
• Audio detection
• Pos info
• Scheduled recording
• Face detection
• Cross line detection
• Intrusion detection

Some of these are marketed features; like alerts when video signals are lost, or when persons cross a boundary defined on Flock’s PTZ cameras. This list is not device-specific^[1], so “Audio detection” is presumably for the Raven scream detector.

As far as this author is aware, “Face detection” is not an advertised feature of any Flock product.

But it’s on the trigger list.

Is it in operational use? If so, by what device(s)?

Contracting government agencies signed away their right to know. And Flock certainly wouldn’t tell us if they did have the feature in active use.

Faceless Identity in 2kB or less

Maybe it doesn’t matter — we know that Flock’s outsourced data labeling platform classifies photos and screams, but that same analysis also revealed that it lets our foreign friends train another feature: ReID. A machine learning algorithm that tracks persons as they move between cameras.

While details on Flock’s actual implementation are sparse, examining an open source project like torch-reid as a model implementation reveals the broader implications.

ReID works by creating a vector, a small (~2kB) file representation of a person. Because of their small size, a virtually unlimited number of these vectors can be stored at extremely minimal cost. The latest 1TB iPhone can hold the vector for every American, Mexican, and Canadian alive, and would still have room for Facebook.

These vectors are not “photos.” They’re very likely also not contractually “Footage.” Whatever rights your PD thinks it has, it probably does not include vectors.

This is combined with soft biometrics, which includes things like clothing (color, type), apparent gender, accessories, and body mechanics (height, body shape, etc.).

That process is much cruder than facial recognition.

The type of data lines up perfectly with the search we’ve seen in moderation logs, where we have someone searching for a “red shirt”:

And with Flock’s advertised person search features:

Where ReID becomes highly accurate (and concerning) is through temporal correlation. Once you know a specific person is at a specific location—maybe because they were seen exiting or, possibly, even driving a vehicle with a license plate tied to a person—you can track them.

No face required.

The quasi-good news is that under laws like Illinois’ Biometric Information Privacy Act, this is likely considered biometric data and its collection is therefore heavily regulated.

Cities with strong protections include Portland, OR, San Francisco, CA, Boston, MA, and New York City. States include Illinois, Texas, and Washington.

The bad news, as always, is that the government does nothing to enforce its own laws.

The End of Public Anonymity

With this, Flock has placed itself into an even more precarious legal situation. Already in trouble after classifying itself as a CJIS compliant entity, now, it’s not only collecting biometrics, it’s tracking people walking.

There is no “driving is a privilege” defense. Existing in public is a fundamental right.

For now, until the government acts, all we can do is wear specific adversarial clothing to defeat ReID and don ski masks to defeat facial recognition.

This is the future your money is buying from Flock.