Jiaotong University training machine through the face recognition criminal accuracy rate of more than 86%

Professor Wu Xilin from Shanghai Jiaotong University and his doctoral student Zhang Xi recently completed a study. They found that through learning, the machine can distinguish who is a criminal and who is a law-abiding citizen through photos. The recognition accuracy rate is above 86%.

This paper is entitled "Automated Inference on Criminality using Face Images" and is currently uploaded on the preprinted website arXiv. They used computer vision and machine learning techniques to detect 1,856 facial photos of Chinese adult men, nearly half of whom were convicted criminals. The experimental results show that through machine learning, the classifier can distinguish the photos of the two groups of criminals and non-criminals with high probability. Especially in the three measures of inner corner spacing, upper lip curvature and nasolabial angle, there is a significant gap between criminals and non-criminals. On average, the criminal's inner corner of the eye is 5.6% shorter than the average person, the upper lip curvature is 23.4%, and the nasolabial angle is 19.6%. At the same time, they found that the difference in facial features between criminals was greater than that of non-criminals.

From ancient times to the present, from the West to the East, we can all see a similar saying that "the heart is born." But whether it is a singer or a psychologist who is engaged in research, he can never get rid of the hat of "superstition" or "discrimination." Wu Yulin and Zhang Xi were out of curiosity and tried to use data analysis to overthrow the ancient "pseudoscience", but the results of the study surprised them. What made them even more unexpected was that once the article was published, it attracted a lot of controversy.

Wu Yulin told the news on November 30 that he received a lot of emails. Although most of the international researchers have sent letters to request data and experimental details for academic exchanges, there are also many unfriendly comments and even accuse him. The research is "irresponsible" to society.

Wu Yulin

"Our luck is not good. When the article just came out, it happened to be before and after Trump was elected. There is an email from the United States saying, 'The United States is now a mess, you should not add chaos.'" Some people directly suggested that Wu Yilin retract the manuscript. . Wu Yulin was somewhat annoyed at the label of “discrimination”. He emphasized that his personal values ​​are absolutely anti-discrimination, and his original purpose in doing this research was to falsify.

In addition, he also received some ridiculous comments, such as some netizens want him to hand this thing to the Discipline Inspection Commission.

Wu Yulin told Yu News that he is still planning to concentrate on this work more rigorously and fully. The maturity of this research is far from the application, and they have no plans to go to the application.

“From another perspective, our research may also provide a basis for anti-discrimination.” But he also admits that how artificial intelligence research should delineate the forbidden zone of value ethics is a very serious issue that is difficult for him to answer.

"There is such a debate in the world now, artificial intelligence has developed to this point."

So, how did Wu Yanlin and Zhang Xi’s research work? Through learning, the accuracy of the machine’s photo identification of criminals is above 86%.

The experiment selected 1,856 photos of Chinese males between the ages of 18 and 55 with no hair occlusion, no scars or other marks on their faces and classified them as criminals and non-criminal groups. The non-criminal group contains 1,126 photos taken from the Internet using “web spiders”. The crowd comes from all walks of life: waiters, construction workers, drivers, doctors, lawyers, professors, etc. The criminal group had a total of 730 photos, of which 330 were from the Ministry of Public Security or the provincial public security bureau, and 400 were provided by a public security bureau that had reached a confidentiality agreement with the experimental group. Of the 730 criminals, 235 involved violent crimes, including murder, rape, physical assault, kidnapping and robbery, while the rest committed non-violent crimes such as theft, fraud and corruption. All photos were sized to 80cm x 80cm and both brightness and gray ratio were controlled to minimize the impact on the results.

A sample of the photos used in the study. Group a is a criminal and group b is a non-criminal.

The experiment used four kinds of classifiers (logical regression, KNN, SVM, CNN) to test the samples and found that they were more successful in classifying offenders and non-criminal groups, with an accuracy rate of over 86%.

The accuracy of the 4 classifiers.

Wu Yulin and Zhang Xi further found that the most significant differences in facial features between criminals and non-criminals were in the three measures of the angular separation of the eyes, the curvature of the upper lip and the angle of the nasolabial angle. On average, the criminal's inner corner of the eye is 5.6% shorter than the average person, the upper lip curvature is 23.4%, and the nasolabial angle is 19.6%.

Figure b shows the three feature points that differ. Table 4 shows the mean and deviation values ​​of the offender group and the non-criminal group at three feature points.

Finally, they found that the “average face” of criminals simulated by non-criminals was similar, but the difference in facial features between criminals was greater than that of non-criminals. That is, the non-criminal groups grow more similar to each other and change less.

Figure c shows the simulated “average face” of the criminal. Figure d shows the “average face” of non-criminals.

"Cranial phase", "born criminals": a study of two centuries of sleep

When many researchers saw the results of Wu Yulin's research, they immediately thought of the theory of "cranial phase" and "born criminals" that were once popular in the 18th and 19th centuries in the West.

In 1870, the Italian prison doctor Long Brosso opened the head of the body of the famous Italian bandit Villera, and found that the head of the skull had a distinct depression, which was positioned like a lower animal. This discovery triggered his inspiration. He proposed the theory of "born sinner", thinking that the criminal is different from the non-criminal in terms of physique, and that the criminal is a phenomenon of returning to the ancestors, with many characteristics of low-level primitives. . At the same time, he believes that crime is hereditary.

The theory of Long Brosso has a great discriminatory color and has been attacked by all parties once published. In addition, because of the lack of data support in the study of Dragon Brosso, it has also been treated as a pseudoscience.

Long Brosso, The Criminal Man.

Since then, the anti-discrimination value ethics has been further developed in Western society, and research on appearance and criminality has gradually become quiet. Until 2011, a psychology research team at Cornell University in the United States found that people only judged whether a criminal was a high success rate by observing a person's photo.

"I also seriously read their research in 2011," Wu said. "But they are taking the test methods of traditional psychology. The MIT Technology Review said that we have taken a new step in this research direction and spoke with data. ”

Originally intended to falsify, the first reaction was very surprised

Regarding the theory of "cranial phase" and "born criminals", Wu Yulin said frankly, "From the perspective of the mainstream scientific community, or from my personal values ​​and personal intuition, I felt that this was unreliable at first. And he believes that the traditional experimental methods, like the psychology of Cornell University, people will inevitably have subjective prejudice and physical fatigue, so he thought of using computers for data analysis. Falsification at the quantitative level.

But Wu Qilin was surprised to see such a result. They achieved results as early as a year ago, but they refused to publish, but repeatedly cross-validated, but never overturned the original conclusion.

For example, in their second edition of the paper uploaded on November 21, they made some changes. As some people suggested that the photos of the criminals were provided by the police, and the cameras used by the police may be different from other cameras in signal, they deliberately added a lot of noise to the optical signals of the photos to drown the difference in the signals of different cameras. However, the previous conclusions are still true, and the classifier still has an accuracy rate of more than 75%. In addition, they specifically re-verified the shooting of the photo of the perpetrator, confirming that it was a common document photo, not taken after being arrested.

“The results from all algorithms are fairly consistent and technically highly reliable.”

"I am here to urge you to withdraw the draft."

The study of the relationship between appearance and criminality seems to be a Pandora's Box. After more than 200 years of sleep, it will be subject to criticism.

Previously, Wu Yilin also showed the media some of the hard-to-find emails he received.

An alumnus of Shanghai Jiaotong University, who is studying abroad, wrote, “I suggest that you revoke this paper and upload an announcement to apologize for inappropriate research methods.” The reason is “this paper is full of extreme discrimination and strong misleading. Researchers in the field of artificial intelligence should not abuse technology and do something that violates ethics."

The letter also highlighted: "This paper has had a very bad influence on the reputation of Shanghai Jiaotong University. This will be a disaster for Shanghai Jiaotong University who apply to American universities."

Another letter came from Cornell University's research colleagues: "I am here to urge you to withdraw the manuscript because it is a shameful job. We can't choose our own lip curvature, eye spacing and so-called nasolabial Angle angle. But the problem of the perpetrator is behavior, not appearance."

There are also some unfriendly voices on the Internet. A user of Hacker News said, "I thought it was a joke when I read the abstract. It turned out to be a serious paper. But this kind of research requires experts in criminology, psychology and machine learning, not just casual. Two people who know Keras."

Some experts in data privacy also pointed out that "the conclusion that such a universality is obtained from such a small data sample will cause great trouble to the innocent people."

According to Dr. Richard Tynan of the International Privacy Protection Organization, “As an individual, you cannot know how the machine gives you a conclusion. On small data sets, algorithms, artificial intelligence and machine learning may establish arbitrary and ridiculous correlations. This is not the fault of the machine. It is dangerous to use complex systems in inappropriate places."

"Our research can also be the basis for anti-discrimination."

Wu Yulin felt annoyed when it was labeled with various unfriendly labels. He said, "My personal values ​​are actually anti-discrimination. For example, some recruitment advertisements openly require good looks, and I am very opposed."

At the same time, he stressed that science is related and cause and effect are two different things. "Criminal tends to have these facial features. It can only be said that there is a correlation between the two. It does not indicate that there is a causal relationship. It is not that the same is a crime, not a natural crime face."

“We just found a correlation between statistical appearance and some social behaviors. We don’t care or discuss the causal logic inside – maybe, maybe not – but we are not experts in this area, no such aspect. Knowledge, training, and experience to do this."

In terms of causality, there may also be an explanation: because some people are different, they may be discriminated against and excluded, and they are more likely to embark on the road of crime. "So our research can also be the basis for anti-discrimination."

“Is nuclear physicist responsible for the damage caused by the atomic bomb?”

But when the news asked how to look at the relationship between social ethical values ​​and scientific research, Wu Yulin admitted that the issue was serious and complicated, and he personally could not answer.

"For the higher interests of human society, is it not for scientists to be self-disciplined, not to say that those who seek truth are not guilty? This controversy has already begun in the world, and artificial intelligence has already reached this point. Is there some? In the restricted area, researchers can't touch it, and frankly I don't know."

“Is nuclear physicist responsible for the damage caused by the atomic bomb?” This is the question that Wu Yulin threw back to the news.

At present, Wu Yulin does not want to be disturbed by some irrational discussions. He is still nervously perfecting the experiment because critics mainly point out that their sample size is not enough. "We use deep learning, and deep learning is also very fashionable now, such as Google's (microblogging) Alpha GO, which requires a lot of sample data. Due to privacy issues, our data is still not big enough, even in published articles. It’s already a big deal, but it’s not enough for deep learning.”

After perfecting, they plan to submit articles to the top international conferences in the field of computer vision, and to the top academic journals such as Nature and Science. But now Wu Yilin is also worried that public opinion will affect the academic acceptance of their research results.

He is looking forward to having more academically rational voices, even if they use data to overthrow their conclusions. Wu Yulin revealed that there is a lot of room for expansion in this field. At present, some international colleagues are doing similar topics. This is why they put arXiv in advance to protect their priorities.

“People’s active attention is good for us, but we cannot misjudge misunderstandings. Our research has not received any support from the private or judicial departments, nor has any commercial purpose.”

Interactive Smart Board

Interactive Smart Board,Smart Boards,Smart Board Projector,Smart Electric Board

ALLIN , https://www.displayapio.com