Facial Recognition Technology and Its Limitations

As it becomes more and more widely adopted, facial recognition technology has many potential applications across several industries, including health, security, marketing and retail, and robotics. Nowadays, facial recognition software is used to scan for individuals’ faces to unlock devices, search for missing children, and track down criminals on the loose. However, while facial recognition offers a wide variety of benefits, it also poses some risks that need to be addressed. New studies suggest that current algorithms are susceptible to bias and inaccurate readings. How will these findings impact facial recognition technology in the future? Let’s find out.

What Is Facial Recognition Technology?

Facial recognition is a form of technology that uses artificial intelligence to match digital images of faces against live footage or pictures of people to identify them. This technology can identify a person based on a single photo and works against the backdrop of other people in a real-time environment. Facial recognition originally came in the form of computer software. But it has also been deployed in commercial applications. That includes smartphones for ID identification and in robotics.

Most computerized facial recognition techniques measure human physical characteristics using mathematical analysis. As such, facial recognition systems are also recognized as a type of biometric identification. Nevertheless, the technology is still in its early stages. It’s also considered less effective than iris identification or even fingerprint recognition. Even then, the contactless and non-invasive processes offered by the technology makes it highly attractive to tech companies.

The Evolution of Facial Recognition Technology

The Evolution of Face Recognition Software — Photo by Arkun Azap via Unsplash

Facial recognition technology has only received media attention relatively recently. But in reality, the technology has been around for some time now. Between the years 2001 and 2004, several government agencies conducted their own research using the algorithm. The algorithm was able to identify faces quickly. However, the rate of accuracy at which the computers identified the faces was still less than the rate at which humans detected and identified faces. Interest in the project would wane for six years. After that, new developments would spark interest in facial recognition.

The breakthrough that they needed finally came in 2010 with the introduction of Convolutional Neural Networks (CNNs). CNNs are a form of machine learning method. The development of CNNs, along with the availability of massive storage spaces and raw processing power, made it possible to develop face recognition technology to a level exceeding human capabilities. To learn more about machine learning, read this ultimate guide about how machine learning works.

How Facial Recognition Works

How Does Facial Recognition Technology Work — Photo by db_oblikovanje via Pixabay

Humans find it easy to encode and recall faces without much effort. We can do it even when the faces in question change with varying facial expressions or age. While facial recognition is a no-brainer task for most people, it can be quite complicated for computers. This might have to do with the fact that whereas humans see faces for what they are, computers see faces as data points.

Most facial recognition technologies use either deep learning or machine learning software. Deep learning allows the program to learn certain objects’ characteristics from raw data. It then applies that knowledge by making decisions and identifying external data. The biggest challenge has to do with identifying real-life, three-dimensional faces that change. And its only basis is two-dimensional images as reference.

Facial recognition technology works in four stages: detection, alignment, feature scanning, and classification. The software first locks in on images of faces from an actual image or an image taken from a video. The facial recognition system will then read the geometry of the faces within the photo. It will then create a facial signature of each one. A facial signature is a combination of facial traits or landmarks that makes a face unique and distinguishable.

The facial recognition software analyzes your facial signature as a mathematical formula. It then compares them with other facial signatures already stored within the database. The computer takes into account measurements of your face. That may include the length of your forehead or the distance between your eyes. Once the software has saved a facial signature, it should be able to recognize that signature from a video or photo.

The Problem With Facial Recognition

The Problems with Facial Recognition — Photo by do Hasbi Saniskoro via Pexels

Few biometric technologies have sparked public interest and imagination quite like facial recognition technology. It has opened up plenty of opportunities for governments and commercial organizations. However, there are also many concerns about privacy, safety, and reliability. The same goes for the upper branches of machine learning. Read this piece to find out about the dangers of the use of artificial intelligence.

Accuracy of Results

The National Institute of Standards and Technology (NIST) examined samples of existing algorithms from several big tech companies. They discovered that most of the systems performed well under ideal conditions. Ideal conditions involve consistent lighting and positioning. They also require that the subjects’ faces are clear and unobscured. But when it comes to unideal conditions where the subject may not be looking directly at the camera or only a portion of the face was captured, the error rates increased significantly.

The software also seems to be unable to take age into account during the analysis. This is because people’s facial signatures can change as they age. This means that if you feed the software a picture of a person taken 20 years ago, chances are that it would fail to match that photo with a more recent picture.

The lack of consistency among existing facial recognition algorithms on the accuracy test is also a problem. Some software brands yield highly accurate results, while the rest yield inaccurate results. But when taken all together, the median accuracy rate was only forty percent. This finding points to the fact that a significant proportion of facial recognition providers on the market have much higher error rates than the leaders in the field. Some providers have managed to create highly accurate software. Others, on the other hand, are still struggling to achieve the same level of reliability.

According to the NIST, the lack of consistency in accuracy levels amongst these providers may be taken as a sign that the software is not ready for commercial production. This also explains why most law enforcement agencies impose confidence thresholds. These force systems to report back only if it can ascertain a degree of confidence in the assessment.

Privacy Issues

Facial recognition technology has blended well into different areas of modern life. But even as we feel comfortable relying on this new technology, there are many issues associated with its use. There are fears that the technology will be used for mass surveillance, racial profiling, and other violations of basic human rights and freedoms. Law enforcement bodies are increasingly reliant on technology. They use it to find potential suspects and witnesses by scanning millions of photos. Security firms also use the software to provide surveillance at public venues like concerts and schools. The software is also a potential tool to commit identity thefts online. Even the commercial industry is increasingly relying on the software to track customer behaviors and movement.

If you feel that your personal data is unsafe, check out these helpful tips to prevent identity theft today.

We already know that facial recognition software needs to be fed images that it can use as a reference to match faces in real-time. But the real question here is whether it is legal or ethically permissible to take photos of individuals and feed them into a database without their consent. This is often the case for law enforcement operations.

The second issue questions the reliability of information collected via facial recognition software. If the technology being used to collect information is unreliable, who’s to say that the information collected is also reliable? Most privacy legislation also prohibits agencies from using the information. That is unless it has been proven to be accurate, up to date, complete, relevant, and not misleading. There is a lack of government regulation, international consensus, or even open discourses about what the technology can and cannot be used for leaves it vulnerable to misuse.

Algorithm Biases

What Causes Bias in Facial Recognition Technology — Photo by Askar Abayev via Pexels

Implications of Algorithm Bias

Researchers have been aware of the susceptibility of facial recognition technology for bias for a long time now. But the issue has deeper implications for society at large. A software that exhibits bias against specific ethnic and racial groups would inevitably replicate the same bias when used in the field. The discussion about software bias is important, particularly in law enforcement applications where individuals’ basic freedoms and rights are at stake. Legal advocates against the use of technology in law enforcement say that such biases would diminish the efficacy of policing and public safety. These biases can discourage people from seeking police assistance. This is especially true for those with previous criminal records.

The growing public concerns prompted a number of tech companies to pull back on the deployment of the software. Meanwhile, the EU mooted a five-year total ban on facial recognition. IBM is one of the companies to abandon facial recognition technology altogether. They informed Congress about this move earlier this year. Amazon and Microsoft also shut down their joint projects, albeit temporarily, until governments have put in place legislation governing the use of the technology. The reform bill Justice in Policing Act approved this year is also intended to prevent racial discrimination in policing by banning the use of facial recognition technology on police officers’ body camera footage.

Many privacy advocates question the ethical and legal permissibility of its application for law enforcement and commercial purposes. Some believe the technology should not be employed until there is consistency amongst software. Or until researchers have made radical improvements in the technology. That entails both accuracy and the reduction or elimination of bias in its decision-making capabilities across demographic variations about race, ethnicity, gender, and other classifications.

What Causes Bias in Facial Recognition

The bias and inaccuracy exhibited by facial recognition algorithms often boil down to how the researchers develop the facial recognition tools. Scientists have identified two potential sources for the bias. The first has to with unequal representation of ethnicities, genders, and social classes in the data sets used to train the algorithms. The second has to do with the tendency of the artificial intelligence component of facial recognition to create biases by mistake.

Scientists need to train algorithms with thousands or even millions of images to allow them to learn what to look for in underactive usage. However, for most algorithms, the majority of the samples used were that of white males. In that case, the system will have a much harder time recognizing other ethnicities or genders. The majority of data sets have more white males in it simply because more photos of them are available on the internet. It comes down to the bias in the data that are fed to the system, which it applies to its decision-making. Scientists believe that researchers can eliminate the inherent biases embedded in the algorithms. But it will require additional training using neutral and representative data sets.

Another concern has to with the inherent imperfections in facial recognition algorithms. These imperfections make the algorithms develop their own biases. We have to remember that these systems are self-learning. That means they make independent inferences about the information being fed to them. The algorithm creates its won models using the data that it can apply to make predictions about what the correct outputs should be for them. However, this is only the ideal scenario. Algorithms sometimes fail to make the right decisions as it uses the wrong parameters for classification, leading to errors for less obvious reasons.

The Future of Facial Recognition Technology

The impact of criticisms on the technology on its future application is not straightforward. While select organizations have abandoned the technology altogether, there are many others who continue to invest incremental resources in the technology. An example would be the security and surveillance industries which owe their success to facial recognition technology. There’s also a waiting list of industries like healthcare, education, and manufacturing seeking to delve into robotics for this purpose.

Other industries are leveraging dimensions of pattern and shape recognition including smart city initiatives. Researchers have also stated that the expected valuation of the technology will surpass $12 billion by the year 2026 compared to the $3.4 billion valuations of the industry in 2019. Companies that continue to support the technology would have to impose their own limitations on the use of technology and face responsibility for its implementation.

There are also a small group of people who are clamoring for the total ban of the technology. But this remains a minuscule proportion as most governments and big tech companies seek a middle ground. This would require the tech industry, governments, and civil society to work together to create a suitable environment for technology without harming citizens’ individual rights. The tech industry needs to improve the accuracy of facial recognition algorithms and remove or eliminate inherent biases. But then again, even if scientists can perfect facial recognition algorithms to achieve perfect accuracy, the technology still faces a myriad of ethical and legal concerns on privacy. This is where the participation of civil society organizations and governments come in. These sectors need to facilitate open discussions and develop federal regulations to ensure the technology’s ethical use across the board.

Final Thoughts on Facial Recognition

There are many ethical and legal issues surrounding facial recognition, but these are not sufficient to completely halt its progress. The algorithms available to us are only still in the infancy stages, and researchers have yet to explore their applications’ full extent. The public must allow researchers to train the algorithms further using representative and neutral data to reduce or eliminate its biases. And on this front, we depend on the tech companies who are keeping up the fight to improve it. But even in this hopeful state, it’s clear that the technology lacks direction and is in urgent need of oversight and control.

Governments and tech companies alike share the responsibility of improving technology and creating measures to ensure its proper use for responsible and ethical use. These reforms shall ultimately determine the future of facial recognition technology or even artificial intelligence in the coming years.