Chinese scientists have taught the neural network to recognize criminals from photographs

Shortly after the invention of the photograph, some criminologists began to notice similar features in the photographs of criminals taken after the arrest. According to their words, criminals are united by common features of a person, according to which they could be attributed to offenders. Modern scientists have tried to prove this theory using the capabilities of artificial intelligence.

An ardent supporter of anthropological theory was the famous Italian criminologist Cesare Lombroso . He believed that the criminals were more than law-abiding citizens, like apes. He was convinced that it was possible to identify monkey features: a sloping forehead, the specific structure of the auricles, various asymmetries of the face and long arms. To prove his point, he took many measurements, although he did not do a statistical analysis of this data.

This omission eventually ruined his theory. English criminologist Charles Goring refuted the views of Lombroso. He analyzed all the information related to the physical disabilities of criminals and law-abiding citizens, and did not find any statistical regularity.
')
Since then, discussions on this topic have subsided until 2011. Then a group of psychologists from Cornell University demonstrated that people are able to distinguish criminals from other people simply by looking at their photos. How was this possible?

Xiaolin Wu and Xi Zhang from Shanghai Transport University tried to answer this question. Scientists used various computer vision algorithms to examine the faces of criminals and law-abiding citizens, and then they checked to see if the machine could tell the difference. They used 1856 photographs of Chinese from the ages of 18 to 55 years. Half of them are criminals. Then scientists took 90% of all photos and taught the convolutional neural network to recognize the difference. The remaining 10% of the images went to the tests.


Faces of criminals


Persons of law-abiding citizens

Wu and Zhang found that a neural network can identify a criminal with an accuracy of 89.5%. “These consistent results are proof of the legitimacy of automated identification of the perpetrator, despite the historical contradictions surrounding this topic,” they say.

According to scientists, there are three facial features by which a neural network identifies a person as a criminal. Compared with people who did not commit offenses, criminals have 23.4% more curvature of the upper lip, the distance from one inner corner of the eye to the other is 6% less, and the angle between the two lines extending from the tip of the nose to the corners of the mouth is 20%. % less.

In their work, researchers demonstrate that these data sets are concentric, but the data of criminal individuals have much stronger deviations. In other words, there is more similarity between the persons of law-abiding citizens, compared with the persons of the criminals. Or criminals have a higher degree of difference in appearance than other people.

Their work explains why the results of some statistical tests make it difficult to see the difference between two sets of data. When Wu and Zhang combined all the portraits of the criminals and all the other portraits to create two "middle" faces, they were almost identical.


"Average" persons: A (criminal) and B (ordinary person), compiled using the Eigenface algorithm; C (offender) and D (ordinary person), compiled by averaging the landmarks and deforming the image

It cannot be said that the results of the work of Chinese scientists were unexpected. If people can cope with this task, it is not surprising that artificial intelligence can do the same. The main question is how people will use these AI capabilities. It is easy to imagine how you can apply the approach of Chinese scientists to data sets like photos from driver's licenses and passports. Thus it is possible to identify those people whom
the machine identifies as probable criminals and then find out if this is really the case.

The paper also says: unlike an expert or a judge, a computer vision algorithm does not have a subjective “baggage in the background,” emotions, prejudices about experience, race, religion, political beliefs, experience. He does not get tired, he does not need sleep or food. It really is. But this does not mean that the machines can not be biased. For example, Beauty.ai was positioned as the first international beauty contest in which participants were evaluated by artificial intelligence. As it turned out later, one of the criteria for evaluation was ethnicity and skin color, for which he was subjected to sharp criticism. The results of the competition showed that the AI ​​preferred the more light-skinned competitors.

Naturally, the work of Chinese scientists needs more serious substantiation and refinement. It is necessary to repeat the experiment with people of different age, gender, ethnic groups and increase the number of data sets. This should help resolve some points of contention. For example, Wu and Zhang believe that criminal faces can be divided into four subgroups, and law-abiding only three. Why it happens? And how will this algorithm work with other groups of people? At the same time, the work raises important questions. If the result really stands up to criticism, then how to explain it? Why do criminal faces have more deviations than ordinary people? How do people define criminals? Is it an innate or acquired skill?

If scientists manage to answer these questions, then perhaps the work of scientists will give a new round in the development of anthropometry of a criminal or other nature.

The scientific work in the latest edition was published on arXiv.org ( ArXiv: 1610.09204 [cs.CV])