A Risk Control System Based On Machine Learning

With the promotion of intelligent terminals,intelligent terminals in life also plays an increasingly important role,such as through the intelligent terminal to complete the payment.Which also brought a huge security problem.For example,an attacker uses stolen equipment to complete the transaction,resulting in damage to the interests of the victim.In order to solve these security problems,the major manufacturers to design a number of traditional account verification can provide more convenient and reliable verification technology,such as through the user's facial features as a user's credentials to authenticate users.Through the facial features to identify users to avoid the cumbersome account password input process,and can be used in non-users when the application of the user is not identified to take a series of measures to prevent users to continue to use the device to prevent the user's loss.However,these third-party applications to collect too much user privacy data,will make users more offensive.So we use the user is relatively insensitive to the motion sensor data as a user feature.This paper designs a set of risk control system based on machine learning,through the study of user behavior habits,training the corresponding model to authenticate the user.The system implements the server-side authentication module and the intelligent terminal protection module,and realizes the real-time protection of the intelligent terminal system.The system adapts to the mainstream Android system and the iOS system,including mobile terminal equipment and intelligent wear equipment.When detecting non-user operation of the device,our system will give the user the appropriate results feedback,and provide a series of automatic protection operations,such as locking equipment and system alarms.At the same time we provide online detection and offline detection of two modes,the equipment only from our cloud to obtain the user's model,you can in the offline environment to authenticate the user to ensure that a variety of environments our system can run.Our work mainly to solve the user data and negative sample imbalance,the user data is not marked the problem,the design of a semi-supervised learning method to achieve the user's authentication.Through the cooperation with the enterprise,collected 1513 users of the sensor data,in our system on the series of data for learning analysis,the results show that the user effective movement and static data accuracy can reach 93.77%and 95.57%.