Researchers believe that with the help of AI, system administrators can better safeguard network security and user rights. Although the current AI can not further predict the specific types of illegal events, it may be able to catch the "invisible" offenders on the network and better protect our security. With the popularity of the Internet, network violations have become a social problem that can not be ignored. The Internet eliminates the space-time distance between potential offenders and victims, so that everyone has objective conditions to break the law, and everyone is in danger of being hurt. The report on characteristics and trends of cybercrime (2016.1-2018.12) released by China judicial big data research institute points out that social platforms, especially QQ and wechat, have become the main tools of virtual crime, through which criminals plan and implement criminal acts on the network. This process does not require real contact, so it is very difficult to capture, which brings many difficulties to law enforcement. Recently, computer researchers from Tokushima University in Japan jointly published a paper on human behavior computing with cyberagent, an agent of a large Japanese network company. They used machine learning methods to analyze the usage data of a social game under cyberagent, and only based on chat times, chat objects Chat time and other basic information can more accurately identify potential network offenders and predict the approximate time of illegal acts. Theoretical basis of "suspect tracking" This is not a whimsical idea. Although we only rely on the Internet to communicate in the game, our online behavior also leaves a large amount of data, which provides rich materials for predicting network violations. The researchers developed this algorithm based on two traditional criminological theories: daily activity theory and social infection theory. The theory of daily activities puts forward that many criminal acts do not occur randomly, and the offender and the victim often intersect in daily activities. For example, in real life, thieves will go to the target site before stealing and observe the behavior law of the target character; Similarly, criminals on the Internet need to contact "prey" in advance to obtain trust. Therefore, there may be a "crime notice" hidden in the player's social activity data. In addition, the theory of social contagion also adds an important point: illegal tendencies or illegal acts can also infect. The most common example is cyber violence. Cyber violence often comes from the wide spread of some passion: under the coercion of groups, some people unconsciously lose their independent judgment ability and inadvertently become online perpetrators. Some studies have pointed out that after "witnessing" the cyber harassment of others in the group, it is also easy for bystanders to attack the same victim or try to harass others. Such infectious behavior also provides important objects and time clues for predicting network illegal events. Based on these two theories, the researchers chose a mobile game called pig party. It focuses on social functions. After logging in, users can dress up virtual rooms and personal images, and communicate with friends or strangers in private, group and public chat. Using the multi-layer nonlinear model, an algorithm good at extracting features from complex data, the researchers analyzed the chat data generated by 550000 users within 6 months, including chat frequency, chat time and message receiver of each user. The heart that wants to do bad things can't escape AI's eyes Researchers combined a variety of neural network models and algorithms to build an artificial intelligence for predicting illegal events. The performance test results show that AI can accurately predict the accounts of future offenders and victims according to user data. Enter the time, frequency and object of the user's chat within two months. The prediction accuracy of AI's illegal account in the next two months can reach 84.85%, and the prediction accuracy of the victim's account is close to 85%. In addition to having a good ability to predict the risk of individual account violations or victimization, AI can basically accurately predict the time of violations in the online community in the next week by providing user activity data within one week. The prediction accuracy of hours and dates is as high as 95.83% and 85.71%, and the results are consistent with the time given by the prediction of victimization. More interestingly, the time when AI alerts the occurrence of illegal events after analyzing the data is not necessarily in the time period of illegal events in the past. It can be seen that it grasps not only the fixed rules, but also the real "logic" in the words and deeds of offenders. AI capable of illegal prediction compresses the massive and scattered records of users' daily activities into data that can be analyzed quantitatively, extracts and understands the laws, and finally forms a strong prediction ability. Researchers believe that with the help of AI, system administrators can better safeguard network security and user rights. Although the current AI can not further predict the specific types of illegal events, it may be able to catch the "invisible" offenders on the network and better protect our security. (◎ sun Linyu, according to global science) (outlook new era)