Facial recognition, which is widespread at airports and Myna Cards, is less accurate for women than for men. “Don’t compromise on the disparity hidden in 99% accuracy” Panasonic’s challenge

The negative aspects of AI, such as “plausible false information” and “hallucinations” created by AI, and “AI bias,” which reproduces social disparities due to bias in learning data such as gender and race, are being closely watched. It is being poured.

The draft guidelines for AI businesses published by the Japanese government list 10 principles as guidelines to consider, including “human-centeredness,” “safety,” and “transparency.”AI’s “fairness”is one of them.

Opportunities to use it are increasing rapidly, such as at airport departure/return gates and My Number card authentication at hospitals and pharmacies.Facial recognition technology is also subject to distortions from AI..

Authentication accuracy is lower for women than for men, and for Asians and other non-whites compared to whites.It’s happening.

This current situation”A human rights issue that cannot be compromised on”Some companies have conducted extensive research on this issue and have come very close to eliminating bias. This is Panasonic Connect, a subsidiary of Panasonic.

Don’t overlook the disparity hidden in 99% accuracy

Panasonic Connect’s facial recognition technology has been introduced in familiar places, such as the departure and return confirmation gates at international airports mentioned above, and My Number card qualification confirmation terminals at medical institutions and pharmacies.

In 2022, the American National Institute of Standards and Technology (NIST) benchmark testError rate 0.206%, or authentication accuracy 99%hit out,Highly rated #1 in the worldHowever, even with such high-precision facial recognition, AI bias occurred.

“There are various ways to collect learning data for facial recognition, but in cases where it is collected from the web, such as academic data, it is easily influenced by the actual world population ratio.

The scale of referenced data differed depending on race and gender, resulting in lower authentication accuracy for specific races and genders.(Kyosei Suzuki, Advanced Sensing Research Department, Advanced Technology Research Institute, Panasonic Connect Technology Research and Development Division)

What on earth does that mean? Please see the diagram above for details. This is a graph that evaluates the accuracy of an existing facial recognition model on an academic dataset.

Depending on the proportion of attributes included in the dataset,Accuracy is lower for non-whites compared to whites and for women compared to men.I understand.

In order to eliminate bias caused by differences in data scale due to these attributes, we have”Thin out the majority”Research has been progressing to find a balance.

However, in some cases, estimating attributes requires manual tagging of images, called annotation.effort and costIt takes.

Also, even if we classify by the attribute of race and adjust the balance,There are many ways to classify gender, age, etc.The problem was that it was difficult to combine and adjust these factors.

The new technology developed by Panasonic Connect isAutomatically determines what attributes should be used to divide data and the classification method (partition), and also finds and learns a model that is common to multiple partitions.This improves the drop in accuracy.

What is the impact of new technology?I tried a simulation at the airport.

Previously, models were trained that were effective at distinguishing facial features from the entire data, so for example, if a facial feature of a Caucasian with a large amount of data was found in the eyes, it would be possible to distinguish between the facial features of a non-Caucasian whose eyes are not a distinctive feature. The focused authentication was being performed.

The new technology first automatically classifies attributes, then derives a model that is effective for distinguishing features for each partition, and then finds and learns a commonly effective model to improve authentication accuracy. .

Developed by Nanyang Technological University in Singapore, it incorporates Invariant Risk Minimization (IRM), which has attracted attention in recent years in the field of image recognition, as a method to remove the effects of data bias, and automatically creates partitions. The company applied a partition learning method called “IP-IRM” to face recognition.

When this was applied to existing face recognition algorithms, an evaluation using an academic dataset showed thatThe average error rate for four races (African, Caucasian, South Asian, East Asian) decreased and authentication accuracy increased..similarlyImprovements in accuracy were also seen when evaluating gender.That’s what it means.

This research was conducted in collaboration with Panasonic Connect, Singapore’s R&D Center Singapore, and Nanyang Technological University, and the paper summarizing the results attracted a lot of attention in 2023.

How much can facial recognition accuracy be improved by using new technology?

According to the paper, out of the 59,890 people of East Asian descent, for whom the data scale is small among the four races, traditionally only33,382 peoplecould be authenticated, but with new technology1246 people increaseWe were able to authenticate 34,628 people.

Similarly, out of a population of approximately 20,000 women, the previous figure was 18,172, but with the new technology,160 people increase18,332 people.

To better understand the impact of improved authentication accuracy through new technology, let’s apply it to the usage history of Haneda Airport’s international passenger terminal. Those who took international flights during the six months from April to September 20239,571,765 peopleOf these, the ones that could be authenticated in the past were7,233,382 peopleIf so, new technology155,063 people increaseof7,388,445 peopleIt is a calculation that can authenticate.