Facial recognition will deliver on its promise if vendors do on theirs, Corsight CEO suggests

Misconceptions around facial recognition and concerns about its effectiveness and how biometric technology will be used are understandable, but must be countered to realize the “force for good” that it can be, according to Corsight CEO Rob Watts.

Watts tells Biometric Update in an interview that he understands these concerns, and where the confusion comes from, attributing a healthy share of the blame to industry players.

“There are lots of companies in this market whose marketing engines are far, far more advanced than their delivery engines,” he states. “I find that abhorrent.”

An example he provides of vendors letting customers and the public down, and a way in which Corsight strives to differentiate itself, is the submission of algorithms which are not commercially available to the NIST FRVT tests which make up the industry’s most famous benchmark. Some submissions from other developers “never see the light of day,” according to Watts. Corsight’s NIST evaluated algorithms are currently in use by customers and available to license.

Truth in advertising, he says, is what will enable facial recognition providers like Corsight to engage in debate, even with libertarians. These difficult conversations are a matter of necessity, Watts suggests, as ethical use and “the whole privacy agenda” are “the biggest bolder in the middle of the road” for industry growth.

Enter the CPO

This is why Corsight recruited former UK Surveillance Camera Commissioner Tony Porter to be the company’s chief privacy officer. That effort took Watts six months.

His job, Watts explains, is to “hold a big stick to” Corsight’s whole operation. To do so, he “kind of sits outside of the business, with good reason.”

Porter, who described his role with Corsight in an interview with Biometric Update in early-2021, is still sought by public sector organizations around the world to lend his expertise to their governance efforts.

With Corsight, Porter works with partners and customers on deployments, and has already blocked near-sales, according to Watts.

“I want our software to be used correctly,” Watts emphasizes. “I want our software to be used for the right reasons.”

Watts is hoping that with greater responsibility from vendors and market education, the conversation will shift, particularly in the U.S., where governance continues to lag behind amid debate over bans.

Ethical responsibility also means ensuring there are not significant differences in algorithm performance when analyzing people from different demographics. Corsight has shown leading results in NIST testing in reducing bias from matching results for Black men compared to white men, Watts notes.

In a follow-up email discussion, Corsight Vice President of Research Ran Vardimon says the company is working to improve its training dataset, with synthetic data offering major potential benefits, and other methods of model-balancing.

The Autonomous AI difference

Watts says that what sets Corsight apart from rest of the market is that “we have the advantage of coming from a different starting point.”

Most – Watts estimates 95 percent — started with an application like a biometric comparison to a passport image. This would occur while the subject is stationary, removes any hats and glasses, and is photographed in a controlled environment. A similar percentage developed their technology based on vectors created by mapping features, he says.

The starting point for Corsight is the use of ‘Autonomous AI’ technology developed by parent company Cortica, similar to the approach to artificial intelligence used in autonomous vehicles.

This results in “huge improvements in terms of speed and accuracy,” according to Watts, to the extent that he was skeptical of Corsight’s claims when he began engaging with the company prior to taking the CEO role.

Corsight claims high accuracy with faces captured while wearing masks, or at a 90-degree head turn.

“We can do it from a drone. I didn’t believe that we could achieve it, but we did it, and can do it.”

“What that means for customers and clients is that they can retain their existing infrastructure,” he adds.

Watts relates an encounter with a police force in the UK which had used facial recognition to investigate a riot, but found no matches from the images of the largely masked crowd. Corsight used the same probe and reference sources to make 147 matches.

“We’re seeing tenders coming out and saying: ‘process 1 million persons of interest 5 seconds,’” he says. “Our system can do that in ten milliseconds.”

This level of accuracy is achieved by the algorithm by the algorithm recognizing faces in a more similar way to how humans do, according to Watts.

Inverted faces are not recognized by feature-based systems, he explains, but like the human eye, Corsight’s can still match them. The algorithm needs 50 pixels between the ears and 50 percent of the face to work, Watts says.

He relates an example of how this kind of matching power can benefit real-world deployments, describing a deployment at a busy port. Corsight’s facial recognition has decreased a processing time to let truck drivers and others enter the secure facility from over an hour to a minute, by matching the face biometrics of visitors while they are driving, is real-time, towards the port at speeds to 30 or 40kmph. The efficiency improvement has been so extreme as to actually increased the country’s GDP in a meaningful way, Watts suggests.

Autonomous AI and transparency

Explaining how exactly results are arrived at remains a challenge for the entire field of autonomous AI, however.

“Being able to achieve real explainability requires full transparency of models being used, such as neural networks, which is even a more general problem (e.g., why do autonomous cars choose a specific action),” Research VP Vardimon says.

“There are interesting ways being assessed about how to achieve explainability, such as highlighting specific features that are unique to the person’s face. This becomes particularly interesting when considering look-alikes, siblings and identical twins. In such cases – what features does the model use tell apart between them and how strong are they? Other questions, such as how can the model’s match confidence reflect image quality, demographic bias, and other conditions (e.g., makeup), are hard to crack.”

While industry and academia continue to work on this challenge, Vardimon says, it is critical for facial recognition to be used correctly as a recommendation tool, rather than a decision-making one.

Vardimon points out that NIST instructs vendors to evaluate bias by reporting false match rate (FMR) at a fixed threshold.

“Corsight’s R&D team dedicates substantial effort to mitigate ethnic and gender bias,” Vardimon writes. “Inspired by NIST views, the main focus of the bias effort is to reach a point where the False Match is the same for different demographic groups. Measuring the target function correctly is essential.

The company has ethically sourced more training data, and is working with GANs on synthetic data. It is working on making its software run efficiently on CPUs, Watts says, and expanding applications “at the hard end of hard.”

Looking ahead

Corsight wants to encourage customers to find new ways to benefit society with facial recognition.

“We’re asking them about their appropriate use,” Watts assures. “We’re asking them about why they’re doing certain things. And equally, we’re asking them the gain they hope to make from having this data.”

Implementations at the network edge and on mobile devices are the main focusses for Corsight right now, Watts says. Personal ownership of biometrics is coming, but not in 2022.

“We’re equally looking at addressing the forensic space as well, so post-event challenges, post-event analytics, much more detail around forensics,” he adds.

Corsight will continue to develop IAM partnerships, and work towards embedding Corsight facial recognition as a unified means of access. “Joe public doesn’t see a divide between cyber and physical,” Watts says, they just want to use the given service or complete the transaction.

This embedded approach means that Corsight’s technology will be used behind-the-scenes in cloud applications, requiring work therefore on anti-spoofing and liveness detection, but end-users will not be aware of who is providing the capability.

With mounting evidence of the benefit of facial recognition, however, and stronger governance, the market can start shifting that boulder out of the road ahead.

Article Topics

accuracy  |  algorithms  |  best practices  |  biometric matching  |  biometric-bias  |  biometrics  |  Corsight  |  Cortica  |  ethics  |  explainability  |  facial recognition  |  marketing  |  NIST  |  privacy  |  research and development