A method for anti-attack in autonomous driving training

Abstract

The present application provides an anti-attack method for autonomous driving training, including: establishing a detection model and image tags based on annotation content; extracting image features to calculate the danger index of the image, wherein the extracting image features to calculate the danger index of the image specifically includes: counting the number of pedestrians and vehicles, identifying the type of road surface on which the vehicle is located in the image, and calculating the danger index of the image; judging whether all the annotators who process the above-mentioned images have abnormal annotation behavior; judging the type of attack by counting the type and number of abnormal behaviors of the annotators; generating the credibility of the annotator account and eliminating low-credibility accounts, wherein the generating the credibility of the annotator account and eliminating low-credibility accounts specifically includes: calculating the credibility of the annotator account and processing the account according to the credibility; calculating the credit score of the account according to the detail and credibility of the annotator account; and allocating annotation tasks to the account based on the account credit score.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to an anti-attack method for autonomous driving training. Background Art

[0002] With the development of science and technology, the automobile industry is undergoing tremendous changes. People's requirements for cars are increasing, and they are pursuing technologies that use science and technology to bring convenience to travel. In particular, the increasing maturity of fields such as artificial intelligence, machine learning, and deep neural networks has made L5 (SAE) level fully automated autonomous driving technology possible in the near future. All driving operations can be completed by the unmanned driving system under all road and environmental conditions, and human drivers can take over driving by manipulating the steering wheel when possible. The process of autonomous driving requires training for marking obstacles, controlling voice, video tracking, etc., but in many cases these tasks are assigned to idle personnel in some labor-intensive areas. Summary of the invention

[0003] The present invention provides an anti-attack method for autonomous driving training, which mainly includes:

[0004] Establish detection models and image tags based on the annotation content; extract image features to calculate the image's danger index; determine whether all annotators who process the above images have abnormal annotation behavior; count the types and times of abnormal behaviors of annotators to determine the type of attack; generate the credibility of the annotator's account and eliminate low-credibility accounts; calculate the credit score of the account based on the detail and credibility of the annotator's account; assign annotation tasks to the account based on the account's credit score;

[0005] Further optionally, the establishment of a detection model and image labels based on the annotation content includes: the annotation content specifically refers to: traffic signs, pedestrians and vehicles, wherein traffic signs specifically include: road traffic lines, traffic lights, signboards, and monitoring, and in the above-mentioned annotation content, pedestrians and vehicles are dangerous factors, and traffic signs are non-dangerous factors; a pre-collected pedestrian image sample set is input as a training set, a pedestrian detection model is trained through a decision tree algorithm, the image to be detected is input into the above-mentioned pedestrian detection model for pedestrian detection, and a pedestrian label of the image is output when a pedestrian exists in the image; a pre-collected vehicle image sample set is input as a training set, and a neural network model is trained through an SVM algorithm to obtain a trained neural network model, vehicle detection is performed on the image to be detected according to the trained neural network model, and the vehicle label of the image is output; traffic sign sample images and non-traffic sign sample images are obtained, a training sample library is established, Haar wavelet features are used to perform feature selection on the images in the training sample library, and the selected features are input into a three-layer perceptron neural network for comprehensive training of traffic signs to obtain a neural network model, wherein the neural network model inputs images and outputs image traffic sign labels.

[0006] Further optionally, the extracting image features to calculate the danger index of the image includes: the counting of the number of pedestrians and vehicles specifically includes:

[0007] Select pictures containing pedestrian labels and vehicle labels as the training set input, train them through a deep convolutional neural network, output the trained target detection model, input the labeled pictures to be detected into the above target detection model, output the detection results and count the number of pedestrians and vehicles monitored in the results.

[0008] The type of road surface on which the vehicle is located in the recognition picture specifically includes:

[0009] Obtain pictures of different road surfaces, and extract road surface features from the pictures, specifically including: using a deep neural network to perform pixel analysis on the pictures, segmenting the ground area from the pictures, extracting road surface features from the ground area using the deep neural network, training a road surface classifier using a naive Bayes algorithm, inputting the pictures to be tested into the road surface classifier, and outputting the road surface type, the road surface type including: asphalt road surface, snow, sand, and mud. The calculation of the danger index of the picture specifically includes:

[0010] Collect safe driving pictures and dangerous driving pictures as training sets. The dangerous driving pictures are pictures of vehicles taken before a traffic accident. The number of pedestrians and vehicles in the training set is output through the above target detection model. The number of pedestrians, the number of vehicles, the ratio of pedestrians to vehicles, and the road type are used as feature values ​​as input. The hazard prediction model is obtained through Logistic regression training. The pictures to be detected are input into the above hazard prediction model to output the hazard index of the picture. The pictures are divided into three hazard levels according to the different hazard indexes: high hazard, medium hazard, and low hazard.

[0011] Further optionally, the judgment and processing of whether all the annotators of the above-mentioned images have abnormal annotation behavior includes: obtaining the account of the annotator of the above-mentioned images with pedestrian labels and vehicle labels and the annotation records within a set time period, and comparing the annotations of the annotator in all the above-mentioned images with the final annotations of the images one by one, and the final annotations refer to the annotation results of the majority of people for the same image, and all inconsistent annotations are obtained and the error type is judged, and the error types include: marking dangerous factors as non-dangerous factors, recorded as error type one, and marking non-dangerous factors as dangerous factors, recorded as error type two. The dangerous factors refer to: pedestrians and vehicles in the annotation content, and the non-dangerous factors refer to: traffic signs in the annotation content. A threshold value for the number of times each type of error is allowed to occur is pre-set. When any type of error exceeds the threshold, it is considered that the annotator has abnormal annotation behavior and his account will be recorded.

[0012] Further optionally, the statistical analysis of the types and times of abnormal behaviors of the annotator to determine the types of attacks includes: statistically analyzing the number of abnormal behaviors in the annotator's account records in a specified time period and the number of error type one and error type two in each abnormal behavior, collecting the annotated records that have been attacked in historical training, and using the annotated records as a training set, specifically including: the number of abnormal behaviors, the number of error type one and the number of error type two, and the types of attacks suffered, the attack types including: malicious attacks, ordinary attacks, and non-attack behaviors, training an attack type prediction model through a linear regression algorithm, inputting the account records to be judged into the above classification model to obtain the types of attacks suffered, and recording and processing the annotator's account accordingly according to different attack types, directly disabling accounts with malicious attack behaviors, and notifying administrators of accounts with ordinary attack behaviors for processing.

[0013] Further optionally, the generating of the credibility of the annotator account and the elimination of low credibility accounts include: The calculating of the credibility of the annotator account specifically includes: Obtaining the historical annotation data of the annotator account within a preset period, obtaining the characteristics of the annotator account according to the historical annotation data, the characteristics of the annotator account including: the working hours within the preset period, the workload within the preset period, normalizing the above characteristic indicators to obtain the normalized characteristic indicator value, performing weighted average processing on the normalized characteristic indicators of the annotator account characteristics to obtain the basic credibility of the annotator, and according to whether the annotator account has abnormal annotation behavior: if there is abnormal annotation behavior, adjusting it on the basis of the basic credibility according to the abnormal annotation behavior, setting different adjustment values ​​according to the determined attack type, and obtaining the credibility of the annotator account within the preset period, otherwise, obtaining the credibility of the annotator account within the preset period according to the basic credibility. The processing of the account according to the credibility specifically includes: Setting a minimum threshold of credibility, when the credibility of the account is lower than the threshold, the account will be deregistered and will no longer participate in the training and annotation process.

[0014] Further optionally, the calculating the credit score of the account according to the detail and credibility of the annotator's account includes:

[0015] The information filled in the annotator's account is obtained to calculate the account detail, where the account detail refers to the ratio of the account's filled in information to all the information that should be filled in, where the information that should be filled in includes: name, gender, date of birth, nationality, valid ID number, education level, contact information, and the number of pictures that can be annotated every day; the weight of each of the above information that should be filled in is determined, the filled in information is recorded as 1, the unfilled information is recorded as 0, and a weighted average is calculated based on the weights to obtain the account detail; the account's credibility is obtained and assigned a first weight value, and the account detail is assigned a second weight value, the first weight value and the second weight value are set according to the specific situation, the weighted average is calculated based on the account's credibility and detail and their corresponding weights, and is recorded as the account's credit score, and its update cycle is determined based on the account's credit score, and the higher the account's credit score, the longer the corresponding credit score cycle. A method for anti-attack training for autonomous driving is characterized in that the system includes: obtaining pictures to be labeled, performing a risk assessment on the content of the pictures through the above-mentioned risk prediction model, dividing the pictures into three groups of high risk, medium risk, and low risk according to the risk level corresponding to the risk index of the obtained pictures, allocating pictures of the corresponding group to the account according to the credit score of the account, obtaining the number of pictures that can be labeled per day filled in the annotator's account, if not filled in, obtaining the number of pictures recorded in the previous marking task of the account, and allocating the corresponding number of pictures in the corresponding group to the account for labeling according to the number of pictures obtained above; if the number of pictures in this group is insufficient, pictures can be allocated to the account from a group with a lower risk level; if there are too many pictures in this group and there are insufficient accounts to allocate, the excess pictures will be classified into a group with a higher risk level for allocation. The technical solution provided by the embodiments of the present invention may include the following beneficial effects: By detecting and restricting the accounts of the labeling personnel, it can be ensured that the accounts that can be used in the system are of good credit, and the regular removal of unused accounts can avoid the waste of system storage space. After the account quality is improved, the accuracy of the autonomous driving training labeling process can be guaranteed, so that the training results can meet the needs of autonomous driving driving. At the same time, it also prevents the training process from being maliciously attacked by criminals from the source, resulting in the inability to conduct training or the training results being greatly affected, thereby avoiding unnecessary waste of time, manpower and financial resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The present invention is a flowchart of an anti-attack method for autonomous driving training. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments. Figure 1 This is a flow chart of an anti-attack method for autonomous driving training according to the present invention. Figure 1As shown, an anti-attack method for autonomous driving training in this embodiment may specifically include:

[0018] Step 101, establish a detection model and image label based on the annotation content. The annotation content specifically refers to: traffic signs, pedestrians and vehicles, wherein traffic signs specifically include: road traffic lines, signal lights, signboards, and monitoring. In the above annotation content, pedestrians and vehicles are dangerous factors, and traffic signs are non-dangerous factors; input the pre-collected pedestrian image sample set as a training set, train the pedestrian detection model through the decision tree algorithm, input the image to be detected into the above pedestrian detection model for pedestrian detection, and output the pedestrian label of the image when there is a pedestrian in the image; input the pre-collected vehicle image sample set as a training set, train the neural network model through the SVM algorithm to obtain a trained neural network model, perform vehicle detection on the image to be detected according to the trained neural network model, and output the vehicle label of the image. Obtain traffic sign sample images and non-traffic sign sample images, establish a training sample library, use Haar wavelet features to select features for the images in the training sample library, input the selected features into the three-layer perceptron neural network for comprehensive training of traffic signs, and obtain a neural network model, the neural network model inputs images and outputs image traffic sign labels.

[0019] Step 102, extract image features and calculate the image's danger index. Count the number of pedestrians and vehicles. Select images containing pedestrian labels and vehicle labels as training set inputs, train them through a deep convolutional neural network, output the trained target detection model, input the labeled images to be detected into the above target detection model, output the detection results and count the number of pedestrians and vehicles monitored in the results. Identify the type of road surface on which the vehicle is located in the image. Obtain images of different road surfaces and extract road surface features from the images, specifically including: performing pixel analysis on the images using a deep neural network, segmenting the ground area from the images, extracting road surface features from the ground area using the deep neural network, training a road surface classifier using a naive Bayes algorithm, inputting the images to be detected into the above road surface classifier, and outputting the road surface type, the road surface types including: asphalt road surface, snow, sand, and mud. Calculate the image's danger index. Collect safe driving pictures and dangerous driving pictures as training sets. The dangerous driving pictures are pictures of vehicles taken before a traffic accident. The number of pedestrians and vehicles in the training set is output through the above target detection model. The number of pedestrians, the number of vehicles, the ratio of pedestrians to vehicles, and the road type are used as feature values ​​as input. The hazard prediction model is obtained through Logistic regression training. The pictures to be detected are input into the above hazard prediction model to output the hazard index of the picture. The pictures are divided into three hazard levels according to the different hazard indexes: high hazard, medium hazard, and low hazard.

[0020] Step 103, determine whether all the annotators who process the above-mentioned pictures have abnormal annotation behavior. Obtain the account of the annotator of the above-mentioned pictures with pedestrian labels and vehicle labels and the annotation records within the set time period, compare the annotations of the annotator in all the above-mentioned pictures with the final annotations of the pictures one by one, the final annotation refers to the annotation results of the majority of people on the same picture, obtain all inconsistent annotations and determine the error type, the error type includes: marking dangerous factors as non-dangerous factors, recorded as error type one, marking non-dangerous factors as dangerous factors, recorded as error type two, the dangerous factors refer to: pedestrians and vehicles in the annotation content, the non-dangerous factors refer to: traffic signs in the annotation content, and pre-set the threshold value of the number of times each type of error is allowed to occur. When any type of error exceeds the threshold, it is considered that the annotator has abnormal annotation behavior and his account will be recorded.

[0021] Step 104: Count the types and times of abnormal behaviors of the annotator to determine the attack type. Count the number of abnormal behaviors in the account records of the annotator in the specified time period and the number of error type 1 and error type 2 in each abnormal behavior, collect the annotated records that have been attacked in historical training, and use the above annotated records as a training set, which specifically includes: the number of abnormal behaviors, the number of error type 1 and the number of error type 2, and the types of attacks suffered. The attack types include: malicious attacks, ordinary attacks, and non-attack behaviors. Train the attack type prediction model through a linear regression algorithm, input the account records to be judged into the above classification model to obtain the types of attacks suffered, and record and process the account of the annotator accordingly according to different attack types. Accounts with malicious attack behaviors are directly disabled, and accounts with ordinary attack behaviors are notified to the administrator for processing.

[0022] Step 105, generate the credibility of the annotator account, and eliminate low credibility accounts. Calculate the credibility of the annotator account. Obtain the historical annotation data of the annotator account within a preset period, obtain the annotator account features according to the historical annotation data, the annotator account features include: working hours within a preset period, workload within a preset period, normalize the above feature indicators to obtain normalized feature indicator values, perform weighted average processing on the normalized feature indicators of the annotator account features, obtain the basic credibility of the annotator, and determine whether the annotator account has abnormal annotation behavior: if there is abnormal annotation behavior, adjust it based on the basic credibility according to the abnormal annotation behavior, set different adjustment values ​​according to the determined attack type, and obtain the credibility of the annotator account within the preset period. Otherwise, obtain the credibility of the annotator account within the preset period according to the basic credibility. Process the account according to the credibility. Set a minimum threshold for credibility. When the credibility of the account is lower than the threshold, the account will be deregistered and will no longer participate in the training and annotation process. Step 106, calculate the credit score of the account according to the detail and credibility of the annotator account. The account details of the annotator are obtained, and the account details are calculated. The account details refer to the ratio of the account information filled in to all the information that should be filled in. The information that should be filled in includes: name, gender, date of birth, nationality, valid ID number, education, contact information, and the number of pictures that can be annotated every day. The weights of the above information that should be filled in are determined, and the filled in information is recorded as 1, and the unfilled information is recorded as 0, and the weighted average is calculated according to the weights to obtain the account details; the credibility of the account is obtained and assigned a first weight value, and the account details are assigned a second weight value. The first weight value and the second weight value are set according to the specific situation. The weighted average is calculated according to the credibility and detail of the account and their corresponding weights and recorded as the credit score of the account, and the update cycle is determined according to the credit score of the account. The higher the credit score of the account, the longer the corresponding credit score cycle.

[0023] Step 107, assigning a labeling task to the account based on the account credit score. Obtain the image to be labeled, evaluate the risk of the image content through the above-mentioned risk prediction model, divide the image into three groups of high risk, medium risk and low risk according to the risk level corresponding to the risk index of the obtained image, assign the image of the corresponding group to the account according to the credit score of the account, obtain the number of images that can be labeled per day filled in the annotator's account, if not filled in, obtain the number of images recorded in the previous labeling task of the account, and assign the corresponding number of images in the corresponding group to the account for labeling according to the above-mentioned number of images. If the number of images in this group is insufficient, the account can be assigned images from the group with a lower risk level. If there are too many images in this group and there are insufficient accounts to be assigned, the excess images are classified into the group with a higher risk level and then assigned. The above description is only an embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention. The program for implementing the information control of the present invention can be written in one or more programming languages ​​or a combination thereof to write a computer program code for performing the operation of the present invention, wherein the programming language includes an object-oriented programming language, such as Java, python, C++, and also includes a conventional procedural programming language, such as C language or a similar programming language. The program code can be executed completely on the user's computer, partially on the user's computer, as an independent software package, partially on the user's computer and partially on the remote computer, or completely on the remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (for example, using an Internet service provider to connect through the Internet). In several embodiments provided by the present invention, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiment described above is only schematic, for example, the division of the unit is only a logical function division, and there may be other division methods in actual implementation. The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, the functional units in the various embodiments of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.The above-mentioned integrated unit can be implemented in the form of hardware or in the form of hardware plus software functional units. The above-mentioned integrated unit implemented in the form of software functional units can be stored in a computer-readable storage medium. The above-mentioned software functional unit is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to perform some steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

Claims

1. A method for preventing attacks during autonomous driving training, characterized in that: The method comprises: establishing a detection model and a picture label based on the annotation content; extracting picture features to calculate the danger index of the picture, wherein the extracting picture features to calculate the danger index of the picture specifically comprises: counting the number of pedestrians and vehicles, identifying the type of road surface on which the vehicle is located in the picture, and calculating the danger index of the picture, wherein the extracting picture features to calculate the danger index of the picture specifically comprises: collecting safe driving pictures and dangerous driving pictures as training sets, wherein the dangerous driving pictures are vehicle driving pictures taken before a traffic accident occurs, outputting the number of pedestrians and vehicles in the training set through a target detection model, taking the number of pedestrians, the number of vehicles, the ratio of pedestrians to vehicles, and the road surface type as feature values ​​as input, obtaining a danger prediction model through Logistic regression training, inputting the picture to be detected into the above-mentioned danger prediction model, and outputting the danger index of the picture; dividing the picture into three danger levels according to different danger indices: high danger, medium danger, and low danger; judging whether all the annotators who process the above-mentioned pictures have abnormal annotation behaviors; judging the attack types by counting the types and times of abnormal behaviors of the annotators; generating the credibility of the annotator account, and eliminating low-credibility accounts, wherein the generating the credibility of the annotator account, and eliminating low-credibility accounts specifically comprises: calculating the credibility of the annotator account, and eliminating low-credibility accounts according to the credibility of the annotator account. The account is processed based on its credibility, specifically including: setting a minimum threshold for credibility. When the credibility of an account is lower than the threshold, the account will be deregistered and will no longer participate in the training and labeling process; calculating the credit score of the account based on the detail and credibility of the annotator's account, and determining its update cycle based on the credit score of the account. The higher the credit score of the account, the longer the credit score update cycle will be; assigning labeling tasks to the account based on the account credit score, including: obtaining the image to be labeled, evaluating the danger of the image content through the above-mentioned danger prediction model, and determining the danger index corresponding to the danger index of the obtained image. The level divides the pictures into three groups: high risk, medium risk, and low risk. According to the credit score of the account, pictures of the corresponding group are allocated to the account. The number of pictures that can be annotated per day filled in the annotator's account is obtained. If it is not filled in, the number of pictures recorded in the previous marking task of the account is obtained. According to the number of pictures obtained above, the corresponding number of pictures in the corresponding group are allocated to the account for annotation. If the number of pictures in this group is insufficient, pictures are allocated to the account from the group of the lower danger level. If there are too many pictures in this group and there are insufficient accounts to allocate, the excess pictures will be classified into the group of the higher danger level for allocation.

2. The method according to claim 1, wherein: The method of establishing a detection model and image labels based on the annotation content includes: the annotation content specifically refers to: traffic signs, pedestrians and vehicles, wherein traffic signs specifically include: road traffic lines, signal lights, signboards, and monitoring. In the above-mentioned annotation content, pedestrians and vehicles are dangerous factors, and traffic signs are non-dangerous factors; a pre-collected pedestrian image sample set is input as a training set, a pedestrian detection model is trained through a decision tree algorithm, an image to be detected is input into the above-mentioned pedestrian detection model for pedestrian detection, and a pedestrian label of the image is output when a pedestrian exists in the image; a pre-collected vehicle image sample set is input as a training set, a neural network model is trained through an SVM algorithm to obtain a trained neural network model, vehicle detection is performed on the image to be detected according to the trained neural network model, and the vehicle label of the image is output; traffic sign sample images and non-traffic sign sample images are obtained, a training sample library is established, Haar wavelet features are used to select features of images in the training sample library, and the selected features are input into a three-layer perceptron neural network for comprehensive training of traffic signs to obtain a neural network model, wherein the neural network model inputs images and outputs image traffic sign labels.

3. The method according to claim 1, wherein: The extracting of image features to calculate the danger index of the image includes: the counting of the number of pedestrians and vehicles, specifically including: selecting images containing pedestrian labels and vehicle labels as training set input, training through a deep convolutional neural network, outputting a target detection model after the training is completed, inputting the labeled images to be detected into the above target detection model, outputting the detection results and counting the number of pedestrians and vehicles monitored in the results; the identifying the type of road surface on which the vehicle is located in the image, specifically including: obtaining images of different road surfaces, extracting road surface features from the images, specifically including: performing pixel analysis on the images using a deep neural network, segmenting the ground area from the images, extracting road surface features from the ground area using the deep neural network, training a road surface classifier through a naive Bayes algorithm, inputting the images to be detected into the above road surface classifier, and outputting the road surface type, the road surface types including: asphalt road surface, snow, sand, and mud.

4. The method according to claim 3, wherein: The judgment and processing of whether all the annotators of the above-mentioned images have abnormal annotation behavior includes: obtaining the annotator account of the above-mentioned images with pedestrian labels and vehicle labels and the annotation records within a set time period, comparing the annotations of the annotator in all the above-mentioned images with the final annotations of the images one by one, the final annotation refers to the annotation results of the majority of people for the same image, obtaining all inconsistent annotations and judging the error type, the error type includes: marking dangerous factors as non-dangerous factors, recorded as error type one, marking non-dangerous factors as dangerous factors, recorded as error type two, the dangerous factors refer to: pedestrians and vehicles in the annotation content, the non-dangerous factors refer to: traffic signs in the annotation content, and pre-setting a threshold value for the number of times each type of error is allowed to occur. When any type of error exceeds the threshold, it is considered that the annotator has abnormal annotation behavior and his account will be recorded.

5. The method according to claim 1, wherein: The method of counting the types and times of abnormal behaviors of the annotator to determine the attack type includes: counting the number of abnormal behaviors in the account records of the annotator in a specified time period and the number of error type one and error type two in each abnormal behavior, collecting the annotated records that have been attacked in historical training, and using the above-mentioned annotated records as a training set, which specifically includes: the number of abnormal behaviors, the number of error type one and the number of error type two, and the types of attacks suffered, wherein the attack types include: malicious attacks, ordinary attacks, and non-attack behaviors, training an attack type prediction model through a linear regression algorithm, inputting the account records to be judged into the attack type prediction model to obtain the types of attacks suffered, and recording and processing the account of the annotator accordingly according to different attack types, directly disabling the account with malicious attack behaviors, and notifying the administrator of the account with ordinary attack behaviors for processing.

6. The method according to claim 1, wherein: The generating of the credibility of the annotator account and the elimination of low-credibility accounts include: the calculating of the credibility of the annotator account specifically includes: obtaining the historical annotation data of the annotator account within a preset period, obtaining the characteristics of the annotator account according to the historical annotation data, the characteristics of the annotator account including: the working hours within the preset period, the workload within the preset period, normalizing the above account characteristics to obtain the normalized characteristic index value, performing weighted average processing on the normalized characteristic index of the annotator account characteristics to obtain the basic credibility of the annotator, and determining whether the annotator account has abnormal annotation behavior: if there is abnormal annotation behavior, adjusting the basic credibility according to the abnormal annotation behavior, setting different adjustment values ​​according to the determined attack type, and obtaining the credibility of the annotator account within the preset period; otherwise, obtaining the credibility of the annotator account within the preset period according to the basic credibility.

7. The method according to claim 1, wherein: The method of calculating the credit score of the account based on the detail and credibility of the annotator's account includes: obtaining the information filled in the annotator's account to calculate the detail of the account, the account detail refers to the ratio of the account's filled in information to all the information that should be filled in, the information that should be filled in includes: name, gender, date of birth, nationality, valid ID number, education level, contact information, and the number of pictures that can be annotated every day; determining the weight of each of the above information that should be filled in, recording the filled in information as 1, and the unfilled information as 0, and calculating the weighted average value according to the weight to obtain the account detail; obtaining the account's credibility and assigning a first weight value, and assigning a second weight value to the account's detail, the first weight value and the second weight value are set according to the specific situation, and the weighted average value is calculated according to the account's credibility and detail and their corresponding weights and recorded as the account's credit score.

Images