Facial recognition and the U.S. Capitol attack

This post examines a number of issues regarding the use of facial recognition. Specifically, it looks at various ways to use facial recognition to identify people who participated in the U.S. Capitol attack.

By TapTheForwardAssist – Own work, CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=98670006

Let’s start with the technological issues before we look at the legal ones. Specifically, we’ll look at three possible ways to construct databases (galleries) to use for facial recognition, and the benefits and drawbacks of each method.

What a facial recognition system does, and what it doesn’t do

The purpose of a one-to-many facial recognition system is to take a facial image (a “probe” image), process it, and compare it to a “gallery” of already-processed facial images. The system then calculates some sort of mathematical likelihood that the probe matches some of the images in the gallery.

That’s it. That’s all the system does, from a technological point of view.

Although outside of the scope of this particular post, I do want to say that a facial recognition system does NOT determine a match. Now the people USING the system could make the decision that one or more of the images in the gallery should be TREATED as a match, based upon mathematical considerations. However, when using a facial recognition system in the United States for criminal purposes, the general procedure is for a trained facial examiner to use his/her expertise to compare the probe image with selected gallery images. This trained examiner will then make a determination, regardless of what the technology says.

But forget about that for now. I want to concentrate on another issue—adding data to the gallery.

Options for creating a facial recognition “gallery”

As I mentioned earlier, the “gallery” is the database against which the submitted facial image (the “probe”) is compared. In a one-to-many comparison, the probe image is compared against all or some of the images in the gallery. (I’m skipping over the “all or some” issue for now.)

So where do you get facial images to put in the gallery?

For purposes of this post, I’m going to describe three sources for gallery images.

  • Government facial images of people who have been convicted of crimes.
  • Government facial images of people who have not necessarily been convicted of crimes, such as people who have been granted driver’s licenses or passports.
  • Publicly available facial images.

Before delving into these three sources of gallery images, I’m going to present a use case. A few of you may recognize it.

Let’s say that there is an important government building located somewhere, and that access to the building is restricted for security reasons. Now let’s say that some people breach that access and illegally enter the building. Things happen, and the people leave. (Again, why they left and weren’t detained immediately is outside the scope of this post.)

Now that a crime has been committed, the question arises—how do you use facial recognition to solve the crime?

A gallery of government criminal facial images

Let’s look at a case in which the images of people who trespassed at the U.S. Capitol…

Whoops, I gave it away! Yes, for those of you who didn’t already figure it out, I’m specifically talking about the people who entered the U.S. Capitol on Wednesday, January 6. (This will NOT be the only appearance of Captain Obvious in this post.)

Anyway, let’s see how the images of people who trespassed at the U.S. Capitol can be compared against a gallery of images of criminals.

From here on in, we need to not only look at technological issues, but also legal issues. Technology does not exist in a vacuum; it can (or at least should) only be used in accordance with the law.

So we have a legal question: can criminal facial images be lawfully used to identify people who have committed crimes?

In most cases, the answer is yes. The primary reason that criminal databases are maintained in the first place is to identify repeat offenders. If someone habitually trespasses into government buildings, the government would obviously like to know when the person trespasses into another government building.

But why did I say “in most cases”? Because there are cases in which a previously-created criminal record can no longer be used.

  • The record is sealed or expunged. This could happen, for example, if a person committed a crime as a juvenile. After some time, the record could be sealed (prohibiting most access) or expunged (removed entirely). If a record is sealed or expunged, then data in the record (including facial images) shouldn’t be available in the gallery.
  • The criminal is pardoned. If someone is pardoned of a crime, then it’s legally the same as if the crime were never committed at all. In that case, the pardoned person’s criminal record may (or may not) be removed from the criminal database. If it is removed, then again the facial image shouldn’t be in the gallery.
  • The crime happened a long time ago. Decades ago, it cost a lot of money to store criminal records, and due to budgetary constraints it wasn’t worthwhile to keep on storing everything. In my corporate career, I’ve encountered a lot of biometric requests for proposal (RFPs) that required conversion of old data to the new biometric system…with the exception of the old stuff. It stands to reason that if the old arrest record from 1960 is never converted to the new system, then that facial image won’t be in the gallery.

So, barring those exceptions, a search of our probe image from the U.S. Capitol could potentially hit against records in the gallery of criminal facial images.

Great, right?

Well, there’s a couple of issues to consider.

First, there are a lot of criminal databases out there. For those who imagine that the FBI, and the CIA, and the BBC, BB King, and Doris Day (yes) have a single massive database with every single criminal record out there…well, they don’t.

  • There are multiple federal criminal databases out there, and it took many years to get two of the major ones (from the FBI and the Department of Homeland Security) to talk to each other.
  • And every state has its own criminal database; some records are submitted to the FBI, and some aren’t.
  • Oh, and there are also local databases. For many years, one of my former employers was the automated fingerprint identification system provider for Bullhead City, Arizona. And there are a lot of Bullhead City-sized databases; one software package, AFIX Tracker (now owned by Aware) has over 500 installations.

So it you want to search criminal databases, you’re going to have to search a bunch of them. Between the multiple federal databases, the state and territory databases, and the local databases, there are hundreds upon hundreds of databases to search. That could take a while.

Which brings us to the second issue, in which we put on our Captain Obvious hat. If a person has never committed a crime, the person’s facial image is NOT in a criminal database. While biometric databases are great at identifying repeat offenders, they’re not so good at identifying first offenders. (They’re great at identifying second offenders, when someone is arrested for a crime and matches against an unidentified biometric record from a previous crime.)

So even if you search all the criminal databases, you’re only going to find the people with previous records. Those who were trespassing at the U.S. Capitol for the first time are completely invisible to a criminal database.

So something else is needed.

A gallery of government non-criminal facial images

Faced with this problem, you may ask yourself (yes), “What if the government had a database of people who hadn’t committed crimes? Could that database be used to identify the people who stormed the U.S. Capitol?”

Well, various governments DO have non-criminal facial databases. The two most obvious examples are the state databases of people who have driver’s licenses or state ID cards, and the federal database of people who have passports.

(This is an opportune time to remind my non-U.S. readers that the United States does not have national ID cards, and any attempt to create a national ID card is fought fiercely.)

I’ll point out the Captain Obvious issue right now: if someone never gets a passport or driver’s license, they’re not going to be in a facial database. This is of course a small subset of the population, but it’s a potential issue.

There’s a much bigger issue regarding the legal ability to use driver’s license photos in criminal investigation. As of 2018, 31 states allowed the practice…which means that 19 didn’t.

So while searches of driver’s license databases offer a good way to identify Capitol trespassers, it’s not perfect either.

A gallery of publicly available facial images

Which brings us to our third way to populate a gallery of facial images to identify Capitol trespassers.

It turns out that governments are not the only people that store facial images. You can find facial images everywhere. My own facial image can be found in countless places, including a page on the Bredemarket website itself.

There are all sorts of sites that post facial images that can be accessible to the public. A few of these sites include Facebook, Google (including YouTube), LinkedIn (part of Microsoft), Twitter, and Venmo. (We’ll return to those companies later.)

In many cases, these image are tied to (non-verified) identities. For example, if you go to my LinkedIn page, you will see an image that purports to be the image of John Bredehoft. But LinkedIn doesn’t know with 100% certainty that this is really an image of John Bredehoft. Perhaps “John Bredehoft” exists, but the posted picture is not that of John Bredehoft. Or perhaps “John Bredehoft” doesn’t exist and is a synthetic identity.

But regardless, there are billions of images out there, tied to billions of purported identities.

What if you could compare the probe images from the U.S. Capitol against a gallery of those billions of images—many more images than held by any government?

It turns out that you CAN perform that comparison, and that law enforcement did perform that comparison.

Clearview AI’s…facial-recognition app has seen a spike in use as police track down the pro-Trump insurgents who descended on the Capitol on Wednesday….

Clearview AI CEO Hoan Ton-That confirmed to Gizmodo that the app saw a 26% jump in search volume on Jan. 7 compared to its usual weekday averages….

Detectives at the Miami Police Department are using Clearview’s tech to identify rioters in images and videos of the attack and forwarding suspect leads to the FBI, per the Times. Earlier this week, the Wall Street Journal reported that an Alabama police department was also employing Clearview’s tech to ID faces in footage and sending potential matches along to federal investigators.

But now we need to return to the legal question: is “publicly available” equivalent to “publicly usable”?

Certain companies, including the aforementioned Facebook, Google (including YouTube), LinkedIn (part of Microsoft), Twitter, and Venmo, maintain that Clearview AI does NOT have permission to use their publicly available data. Not because of government laws, but because of the companies’ own policies. Here’s what two of the companies said about a year ago:

“Scraping people’s information violates our policies, which is why we’ve demanded that Clearview stop accessing or using information from Facebook or Instagram,” Facebook’s spokesperson told Business Insider….

“YouTube’s Terms of Service explicitly forbid collecting data that can be used to identify a person. Clearview has publicly admitted to doing exactly that, and in response, we sent them a cease-and-desist letter.”

For its part, Clearview AI maintains that its First Amendment government rights supersede the terms of service of the companies.

But other things come in play in addition to terms of service. Lawsuits filed in 2020 allege that Clearview AI’s practices violate the California Consumer Privacy Act of 2018, and the even more stringent Illinois Biometric Information Privacy Act of 2008. BIPA is so stringent that even Google is affected by it; as I’ve previously noted, Google’s Nest Hello Video Doorbell’s “familiar face” alerts is not available in Illinois.

Between corporate complaints and aggrieved citizens, the jury is literally still out on Clearview AI’s business model. So while it may work technologically, it may not work legally.

And one more thing

Of course, people are asking themselves, why do we even need to use facial recognition at all? After all, some of the trespassers actually filmed themselves trespassing. And when people see the widely-distributed pictures of the trespassers, they can be identified without using facial recognition.

Yes, to a point.

While it seems intuitive that eyewitnesses can easily identify people in photos, it turns out that such identifications can be unreliable. As the California Innocence Project reminds us:

One of the main causes of wrongful convictions is eyewitness misidentifications. Despite a high rate of error (as many as 1 in 4 stranger eyewitness identifications are wrong), eyewitness identifications are considered some of the most powerful evidence against a suspect.

The California Innocence Project then provides an example of a case in which someone was inaccurately identified due to an eyewitness misidentification. Correction: it provided 11 examples, including ones in which the witnesses were presented to the viewer in a controlled environment (six-pack lineups, similar backgrounds).

The FBI project, in which people look at images captured from the U.S. Capitol itself, is NOT a controlled environment.

Leave a Comment

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s