- What methods are used for facial recognition in public surveillance? Ideally, an answer would point to the software, algorithms or specifications being used.
- How can those be fooled?
- Fake or disfigured face? To what extend would one need to employ artificial scars or moles, make-up or even a complete face mask?
- Will spectacles work? Sun glasses vs. normal ones?
- Will using a hat to hide a portion of a face work? How much needs to be hidden? Hair, forehead, eyes, nose or complete face?
- Blinding (not: destroying) a camera with laser or LEDS, provided the camera can be seen, and one can aim at it?
- What else?
- Slightly related: Are there any other methods being used, such as clothes detection, gait detection, etc?
Appropriate Question if Answered in Brief
A shelf of books may be required to answer this question comprehensively, however a brief answer may be useful to the community. It is, thus far, customary for the allowable breadth of question in this stack exchange to be much greater than that of stack overflow. This is certainly appropriate because of the breadth of artificial intelligence as a largely interdisciplinary science and technical field.
How good is facial recognition employed in public surveillance? Let's assume that by, "good," is meant few false negatives and few false positives. To keep it simple for this brief answer, we can aggregate the quality of the recognition of human identity as a scalar of this two dimensional vector [Pfp, Pfn], where Pfp is the probability of a false positive and Pfn is the probability of a false negative, as follows.
q = [ (1 - Pfp2) * (1 - Pfn2) ]8
With this simple formula we have the desirable aggregation behavior of intolerance of poor quality in either dimension.
The most limiting factors in driving quality
q to 1.0 is neither face recognition algorithm nor face recognition software nor optics. The limiting factors are these.
- Finding, tracking, and zooming to faces with moving or fixed obstructions or reflective surfaces, such as windshields
- The cost of the computing power required in places like malls and urban sidewalks
Specific Technology and Specifications
The emerging field of facial recognition is a subject of much research and the best algorithms and software are the critical intellectual property of the largest corporations in the world, however you can look up what public facing work is being done at the university and public levels by searching through scholarly articles using the simple search term, "successful facial recognition deep learning."
You will need to read many and look up terms for a few weeks before you will know what they are talking about. A specialized nomenclature and terminology has formed.
The specification being used is the one above. The goal is to uniquely identify a person better than a social security number, using skin contour feature recognition, which is based on bone and cartilage. Such technology will likely never be as reliable as fingerprints or DNA gene sequencing but is much easier to sample.
Defeating Tracking and Countermeasures
Any software solution that attempts to extract contour features based on skeletal geometry, cartilage, or fat thickness patterns is vulnerable to what security professionals call spoofing. In the realm of game theory, this is called a countermeasure.
Masking is an obvious method to defeat facial recognition but very limited in use, since people don't normally mask their face. They do wear visors, glasses, and sun glasses, so that can impede reliability in facial recognition. Surgery is a more effective countermeasure but requires a sacrifice in money, time, and socially, and can also draw attention.
Prosthetic facial devices are available but expensive and can be detected because side views reveal a distance change of more than three standard deviations between the other features of the head and the front surface of the prosthetic.
Clothes detection has obvious drawbacks. Gait detection is problematic too, but the detection of the centers of rotation in the skeletal system is actually much more reliable than facial recognition. Theoretically, precise measurements between the centers of rotation of all major joints can be detected from one or more views of a person walking, talking, writing, and playing sports. This is as identifying as a face under certain conditions.
How far various intelligence agencies and corporations have progressed along these lines is top secret and company confidential respectively. If I have worked on any visual 3D reconstructive feature extraction for the purpose of human identification I will have deliberately omitted such technologies and inventions from the above paragraphs for legal and national security reasons.
Defeating recognition by radio or radiant emissions to exploit optical or transmission vulnerabilities is possible but probably illegal in most jurisdictions and as likely to draw attention as entering a bank wearing a Halloween mask.
Much of this is moot for the average person who drives around in a car with a registration plate that is much easier to acquire and uniquely identify and keeps a mobile device connected with a public cell network. They communicate via unsecured emails and FaceBook their meals, cats, and family. Yet average people are not usually vulnerable in any meaningful way because they are of no interest.
From a privacy point of view, there is nothing more effective, as easy, and as inexpensive as mediocrity. If a person breaks only the laws others are breaking (marginally speeding and stealing pens), acts socially normal, has a typical opinion about politics, and engages in the discussion of the same redundant topics as does the general public, no one will likely ever look for that person's face. No one will be likely to ever track a comprehensively unexceptional person other than as a single data point inside some gigantic demographic category to market to.
Privacy and Accomplishment
If a person wishes to do something exceptional in complete privacy, they can develop it inside a private space they own, with no visibility from outside and no visitors ever. They can then release their great achievement to everyone they can find who will understand its value just before they die.
Such a lifelong scheme requires intense and sustained self-discipline. Only a few great people in human history have pulled it off. Most great things are accomplished in some level of public view and with the substantial sacrifice of privacy. Risk for the average exceptional person is mitigated most often by hiring security.
Large scale IT projects from the government have a high probability to fail. Such projects are usually run out of budget (which costs at least 1 billion US$), aren't reaching the goals and result into mismanagement. If somebody is trying to inject Artificial Intelligence in such IT projects, the failure rate is much higher. In contrast to a normal software project which is based on well understand technology, Artificial Intelligence is a new topic and has no proven outcome. Video surveillance with multi-cameras and deeplearning hardware can be called without any doubt an Artificial Intelligence application. That means, it is 99% sure, that at the end of 10 years of implementation, the nvidia hardware isn't working, the Python scripts fail, and a simple sun glass will irritate the surveillance camera too much, so it won't recognize anything.
Instead of trying to use modern AI technology which isn't tested, it is often the better idea to prefer low-tech solution. That means, an analog CCTV camera plus a human in front of the screen will reach better results then a fully autonomous system with cloud based image recognition algorithms. Instead what ongoing campaigns like Smartcity or Skymind are saying, the technology isn't ready for practical usage, will produce high costs and can recognize nobody.