At Rise of AI 2019 in Berlin, Shaun Moore, CEO of Trueface, argues that facial recognition is becoming a mainstream, real-time technology as camera infrastructure scales and processing shrinks from hundreds of megabytes to a few megabytes that can run at the edge on devices like a Raspberry Pi. He traces the technology from hand-drawn proportion analysis in 1960 through the failed 2002 Super Bowl trial, Facebook's 2010 auto-tagging research, and Google's 2015 false claim of 100 percent accuracy, which was undermined by its gorilla-tagging error. Moore describes applications in travel, school safety, fraud reduction, access control and convenience, while emphasizing that public training datasets are skewed 70-80 percent white male, producing biased algorithms unless companies source balanced or synthetic data, including data generated with generative adversarial networks. He distinguishes anonymous, template-based recognition from data storage that identifies individuals, and calls for federal or international regulation, arguing that responsible use, not banning the technology, is the right path forward.
Key takeaways
- Facial recognition models have shrunk from 400-500 megabytes to about 5 megabytes, letting them run locally on edge devices such as a Raspberry Pi.
- Google publicly claimed 100 percent face-match accuracy in 2015 but its software mistakenly tagged people as gorillas shortly after.
- Anonymous facial recognition can match faces against a watchlist without storing names or other personal data, addressing a key privacy objection.
- Training datasets that are 70-80 percent white male produce biased algorithms; GAN-generated synthetic faces can help correct that skew.
- Facial recognition cut boarding time for a 777 flight from LAX to Heathrow from 50 minutes to 25 minutes.
Detailed notes
What facial recognition is and how it developed
Moore defines facial recognition as a mathematical representation of a face: for Trueface, 512 characteristics of facial landmark proportions that differ by vendor. He dates the field to about 1960, when Woodrow Wilson Bledsoe hand-drew proportion comparisons between faces, with computers taking over the calculations by 1991. The 2002 Super Bowl in Tampa was the first major US commercial deployment, used to search for criminals; it produced many false positives and was shut down within three months. Facebook began researching auto-tagging with facial recognition in 2010 and by 2014 claimed 97.6 percent accuracy, matching human-level performance. In 2015 Google claimed 100 percent accuracy but was shown to be wrong when its software tagged people as gorillas. The iPhone X in 2017 made facial recognition a mainstream consumer technology. Moore notes about 650 million cameras exist worldwide today, rising to roughly 1 billion within a year and a half, and that recognition models have shrunk from 400-500 megabytes to about 5 megabytes, letting them run locally on devices such as a Raspberry Pi, as already used in smart video doorbells.
Anonymous recognition and privacy concerns
Describing a deployment at Oktoberfest in Munich, where security relied on 27 human "super recognizers" to monitor six million visitors, Moore argues human monitoring alone is not effective at that scale. He presents anonymous facial recognition, in which a system matches faces against a watchlist without storing names, ages or other personal data, as a way to avoid the main privacy objection: uncertainty about where facial data is stored and who can access it. He raises the open question of who owns the mathematical representation generated from a person's face once a company like Trueface has created it. He says he does not believe mass surveillance states are likely in the US or Europe, because facial recognition companies and governments there lack the capacity, but that bias and personal privacy are the more pressing concerns.
Bias in training data
Moore returns to the 2015 Google gorilla-tagging failure as evidence that facial recognition can misidentify people of color, and cites a 2018 MIT Media Lab study concluding that governments should not use facial recognition because bias causes it to disproportionately target non-white, non-male individuals. He attributes this to publicly available training datasets that are 70-80 percent white male. Data can come from public sources such as IMDB, from private data supplied by clients, or from synthetic data generated with generative adversarial networks, which Moore demonstrates by showing an audience images of faces that, he says, do not belong to any real person. He argues GAN-generated data can be customized for a target demographic at scale and low cost, reducing bias in training sets.
Applications across travel, security and commerce
Moore lists concrete use cases: at LAX to Heathrow, facial recognition cut boarding time for a Boeing 777 from 50 to 25 minutes; school security systems can blur and immediately discard footage of students unless a known threat is matched; banks are investing in facial recognition to cut ATM, desktop and retail fraud; access control at construction sites and office buildings can cut a 45-minute morning entry rush to about 15 minutes; and smart doorbells identify visitors and can play audio to known guests, useful for scenarios like Airbnb access. He says Trueface's system can scan a billion faces in one second, and distinguishes one-to-one face matching (unlocking a phone) from one-to-many matching (finding a person in a crowd).
Regulation and the future of the technology
Moore welcomes regulation, citing a California law, passed shortly before the talk, that requires police to obtain judicial approval before using facial recognition, and argues for federal or international rules, comparing how differently defense agencies, phone makers and social platforms deploy the technology. He forecasts 3D facial recognition becoming practical in five to ten years as sensor infrastructure shifts from 2D, and real-time personalized retail advertising, opt-in only, arriving in about five years. He notes age estimation from a face already achieves a mean average error of 2.5 years. He closes by saying the technology will only see mass adoption if it is used responsibly, with checks and balances in place.
Q&A highlights
Asked about accidental consent, such as someone scanning a sleeping person's face for payment, Moore says liveness checks such as heartbeat sensing through the forehead and micro-movement or blink detection over 15-30 frames guard against this, though he notes Apple has had public trouble in this area. Asked about dependence on responsible governments, he agrees ethical regulation is needed and says the guiding principle should be that humans, not algorithms, make final decisions affecting human lives, contrasting this with heavy state investment in facial recognition in China. On twins, he says there is no government study to cite, but that Trueface has correctly distinguished every twin pair in roughly 15 to 20 informally tested cases. On adversarial attacks, such as clothing patterns or makeup designed to evade recognition, he calls it a permanent "cat and mouse" problem and says he supports regulating such evasion attempts. Asked whether stolen model parameters could be used with generative networks to reverse-engineer a matching face, he says this would require insider access to Trueface's source code and that encryption limits the risk.
Quotes
If this technology is used responsibly, we're very confident that we can provide a safer and smarter world for everyone.
Shaun Moore
The most important thing is that humanity comes first, there's total transparency.
Shaun Moore
It's a game of cat and mouse and it always will be.
Shaun Moore
Questions answered in this talk
How does facial recognition technology work?
According to Shaun Moore at Rise of AI 2019, facial recognition creates a mathematical representation of a face using facial landmark proportions, 512 characteristics in Trueface's system, that is vendor-specific and cannot be reverse-engineered to reconstruct the original face.
Is facial recognition biased against people of color?
According to Shaun Moore at Rise of AI 2019, publicly available training datasets used by many facial recognition vendors are 70-80 percent white male, which produces biased algorithms that disproportionately misidentify people of color, a finding also reported by a 2018 MIT Media Lab study.
What happens if a facial recognition template is stolen or compromised?
According to Shaun Moore at Rise of AI 2019, a compromised facial template can simply be deleted and a new 512-digit representation issued, similar to resetting a password, because the template is a mathematical representation rather than the face itself.