Facial Recognition Technology will have a severe impact on society in the future, resulting in a loss of anonymity for everybody. The Russian artist Egor Tsvetkov demonstrated with an experiment that it is already possible today to identify random people on the metro by just using their photos and a facial recognition app. Currently this was done by processing images afterwards with face search and augmenting it with human verification, but it is not difficult to imagine every cell phone or connected car being able to identify anyone’s face and perform a search on the web within less than a second. Instead of classical fears of big brother’s mass surveillance, there might be technology ready for peer-to-peer mass surveillance.
The progress of image recognition
A decade ago the most sophisticated computers could not tell a dog from a cat. With some funny exceptions aside (Chihuahua or a muffin? Labradoodle or fried chicken? Sheepdog or a mop?), image recognition algorithms have surpassed human performance. According to the Electronic Frontier Foundation’s AI Progress Measurement project, recent progress in image recognition on the ImageNet benchmark has outperformed humans several times last year. Combining that progress in precision with Moore’s law of growing computational power, one can expect similar progress in the time it takes to correctly identify images.
A recent report on the The Malicious Use of Artificial Intelligence considers facial recognition systems to be both efficient and scalable: ‘Once [facial recognition system] is developed and trained, it can be applied to many different camera feeds for much less than the cost of hiring human analysts to do the equivalent work.’
Already, there are examples of facial recognition algorithms being used. Facial recognition sunglasses used by Chinese police are currently able to find and match faces of 10,000 suspects in 100 milliseconds. Fluctuating accuracy, false-positives and a small search database do not yet allow scaled- up and ubiquitous implementation. But, the continuation of technological progress suggests that several orders of magnitude larger databases searchable within the same period of time will be a reality in a few years.
Going undercover: Adversarial patches and eye patches
The most obvious reaction to facial recognition – covering one’s face – might not help, since the next frontier for facial recognition is identifying people whose faces are covered. One trend in reaction to ubiquitous facial recognition technology is exploring how fashion can be used as camouflage from face-detection technology. There is a renewed interest in finding ways to use make-up and fashion items, for instance eye patches, to make the face unrecognizable to computers.
The other trend is to reverse engineer fundamental working principles of facial recognition and use artificial intelligence to find ways to derail their detection mechanism. This has lead to the development of adversarial patches and one-pixel attack vectors custom made to affect facial recognition technology. However, protecting one’s privacy in such a way might have unintended consequences: self-driving car software might not be able to classify pedestrians correctly in the street and smart industrial machinery and robots might not be able to detect workers in factories, resulting in potentially fatal tragedies.
Facial recognition and the clash of regulations
Neither hardware nor software responses to facial recognition might ultimately prove enough. If we extrapolate from the trend that software can already automatically extract 52 data points aboutpedestrians, including breathing and heart rate, it will become increasingly difficult to hide behind camouflage. At the same time, evading detection through reverse-engineering recognition might have its own unintended consequences. Therefore, facial recognition and image recognition in general urgently need a regulatory response.
However, recent regulations on covering one’s face are going in the direction of requiring less anonymity. Belgium, France and several other European countries have recently introduced legislations partially or completely banning full face veils. “Anyone who wears a garment that hides the face in public will be punished with a fine,” says the new Danish law. In the case of Belcacemi and Oussar v. Belgium the European Court of Human Rights emphasized that ‘the ban was justifiable in principle, solely to the extent that it sought to guarantee the conditions of “living together”‘ and that ‘it was a matter of protecting a condition of interaction between individuals which, for the State, was essential to ensure the functioning of a democratic society.‘
The accelerating development facial recognition technology and its potential impact on society were completely ignored in those discussions. This is unfortunate, and potentially dangerous when we add the new puzzle piece of facial recognition to the equation. If facial recognition technology becomes so ubiquitous as to disrupt social life, it would be rather depressive to suddenly find ourselves to realize that we have recently banned all means to protect our privacy from mass crowd-surveillance.