A training method and device of a risk judgment model for face recognition

Abstract

The embodiment of the specification provides a training method and device of a risk judgment model for face recognition, the method comprises the following steps: obtaining a first sample, the sample comprising an identification of a first user, an identification of a first device and a first label value representing whether recognition is correct or not; performing graph convolution operation on a user similarity graph through a first graph convolution network to obtain a first user vector of the first user; performing graph convolution operation on a user-device relationship graph through a second graph convolution network to obtain a second user vector of the first user and a first device vector of the first device; performing graph convolution operation on a device relationship graph through a third graph convolution network to obtain a second device vector of the first device; combining the first and second user vectors and the first and second device vectors respectively to obtain a third user vector and a third device vector; and updating the first, second and third graph convolution networks with the purpose that a dot product value of the third user vector and the third device vector tends to the first label value.

Description

TECHNICAL FIELD

[0001] One or more embodiments of the present specification relate to the field of image recognition and risk prediction, in particular to a training method and device of a risk judgment model for face recognition. BACKGROUND

[0002] At present, with the wide application of face recognition technology in different industries and different application scenarios, the equipment for face recognition has rapidly developed from personal terminal equipment mainly used for online face recognition scenarios to public recognition equipment mainly used for offline face recognition scenarios. Due to the huge number of users facing the public recognition equipment compared with the personal terminal equipment, the difficulty of face recognition is increased, which significantly increases the risk of face recognition on the public recognition equipment.

[0003] Therefore, there is a need for a new risk judgment method for face recognition on public equipment. SUMMARY

[0004] Embodiments of the present specification aim to provide a training method and device of a risk judgment model for face recognition. The risk judgment model trained by the method can quickly determine the risk degree of face recognition on public equipment, and then can improve the success rate of face recognition on public recognition equipment according to the risk degree, solving the deficiencies in the prior art.

[0005] According to a first aspect, a training method of a risk judgment model for face recognition is provided, the risk judgment model comprising at least a first, second and third graph convolutional network, the method comprising:

[0006] obtaining a first sample, the first sample comprising an identification of a first user, an identification of a first equipment, and a first label value, the identification of the first user being a first recognition result output by the first equipment for face recognition, the first label value being used to indicate whether the first recognition result is an incorrect recognition result;

[0007] performing graph convolution operation on a pre-constructed user similarity relation graph by the first graph convolutional network to obtain a first user vector corresponding to the first user; the user similarity relation graph is constructed according to a first similarity between a plurality of users, the first similarity being determined based on face images of the plurality of users, the plurality of users including the first user;

[0008] performing graph convolution operation on a pre-constructed user-device relationship graph by the second graph convolution network to obtain a second user vector corresponding to the first user and a first device vector corresponding to the first device; the user-device relationship graph is constructed according to historical association information between the plurality of users and a plurality of public face recognition devices, and the plurality of public face recognition devices include the first device;

[0009] performing graph convolution operation on a pre-constructed device relationship graph by the third graph convolution network to obtain a second device vector corresponding to the first device; the device relationship graph is constructed based on geographical positions of the plurality of public face recognition devices;

[0010] combining the first user vector and the second user vector to obtain a third user vector corresponding to the first user; and combining the first device vector and the second device vector to obtain a third device vector corresponding to the first device;

[0011] updating network parameters of the first, second and third graph convolution networks so that a dot product value between the third user vector and the third device vector tends to the first label value.

[0012] In a possible implementation, the combining the first user vector and the second user vector to obtain the third user vector corresponding to the first user includes:

[0013] obtaining values of each dimension of the third user vector according to a superimposed value / average value of values of each dimension of the first user vector and the second user vector.

[0014] In a possible implementation, the superimposed value of the values of each dimension of the first user vector and the second user vector includes a weighted superimposed value of the values of each dimension of the first user vector and the second user vector.

[0015] In a possible implementation, the risk judgment model further includes a first attention network.

[0016] the weighted weights of the values of each dimension of the first user vector and the second user vector in the weighted superimposed value are determined based on the first attention network.

[0017] The updating the network parameters of the first, second and third graph convolution networks so that the dot product value between the third user vector and the third device vector tends to the first label value includes:

[0018] updating network parameters of the first, second, third graph convolution networks and the first attention network so that the dot product value between the third user vector and the third device vector tends to the first label value.

[0019] In a possible implementation, the combining the first device vector and the second device vector to obtain the third device vector corresponding to the first device comprises:

[0020] The value of each dimension of the third device vector is obtained according to a superimposed value / average value of the values of each dimension of the first device vector and the second device vector.

[0021] In a possible implementation, the superimposed value of the values of each dimension of the first device vector and the second device vector comprises a weighted superimposed value of the values of each dimension of the first device vector and the second device vector.

[0022] In a possible implementation, the risk judgment model further comprises a second attention network.

[0023] The weighted weights of the values of each dimension of the first device vector and the second device vector in the weighted superimposed value are determined based on the second attention network.

[0024] The updating the network parameters of the first, second and third graph convolution networks so that the dot product value between the third user vector and the third device vector tends to the first label value comprises:

[0025] The updating the network parameters of the first, second, third graph convolution networks and the second attention network so that the dot product value between the third user vector and the third device vector tends to the first label value.

[0026] In a possible implementation, the user similarity relation graph comprises a plurality of user nodes and edges between the user nodes, the plurality of user nodes respectively correspond to the plurality of users, and the edges correspond to first similarities between the users.

[0027] In a possible implementation, the user-device relation graph comprises a plurality of user nodes, a plurality of device nodes and edges between the user nodes and the device nodes, the plurality of user nodes respectively correspond to the plurality of users, the plurality of device nodes respectively correspond to the plurality of public facial recognition devices, and the edges correspond to historical association information between the users and the public facial recognition devices.

[0028] In a possible implementation, the device relation graph comprises a plurality of device nodes and edges between the plurality of device nodes, the plurality of device nodes respectively correspond to the plurality of public facial recognition devices, and the edges correspond to first location relations between the public facial recognition devices.

[0029] In a possible implementation, the first position relationship is used to indicate that the distance between the public face recognition devices is less than a first threshold.

[0030] In a possible implementation, the historical association information between the user and the public face recognition device includes one or more of face swiping, authentication, and payment behaviors of the user through the public face recognition device.

[0031] In a possible implementation, the user similarity relationship graph is constructed according to the first similarity between the plurality of users.

[0032] In a possible implementation, the plurality of users further includes a second user, and the plurality of public face recognition devices further includes a second device, and the method further includes:

[0033] obtaining a fourth user vector of the second user by performing graph convolution operation on the user similarity relationship graph by the updated first graph convolution network;

[0034] obtaining a fifth user vector of the second user by performing graph convolution operation on the user and device relationship graph by the updated second graph convolution network;

[0035] combining the fourth user vector and the fifth user vector to obtain a sixth user vector corresponding to the second user.

[0036] In a possible implementation, the method further includes:

[0037] obtaining a fourth device vector of the second device by performing graph convolution operation on the user and device relationship graph by the updated second graph convolution network;

[0038] obtaining a fifth device vector of the second device by performing graph convolution operation on the device relationship graph by the updated third graph convolution network;

[0039] combining the fourth device vector and the fifth device vector to obtain a sixth user vector corresponding to the second device.

[0040] In a possible implementation, the method further includes:

[0041] obtaining a second recognition result output by the second device for face recognition, the second recognition result including an identity of the second user;

[0042] According to the identity of the second user and the identity of the second device, a sixth user vector corresponding to a second user and a sixth device vector corresponding to the second device are determined;

[0043] According to the sixth user vector and the sixth device vector, a probability that the second recognition result is an error recognition result is determined.

[0044] According to a second aspect, a device for training a risk judgment model for face recognition is provided, the risk judgment model comprising at least a first, a second and a third graph convolution network, the device comprising:

[0045] a sample obtaining unit configured to obtain a first sample, the first sample comprising an identity of a first user, an identity of a first device, and a first label value, the identity of the first user being a first recognition result output by the first device through face recognition, the first label value being used to indicate whether the first recognition result is an error recognition result;

[0046] a first graph convolution unit configured to perform graph convolution operation on a pre-constructed user similarity graph through the first graph convolution network to obtain a first user vector corresponding to the first user, the user similarity graph being constructed according to a first similarity between a plurality of users, the first similarity being determined based on face images of the plurality of users, the plurality of users including the first user;

[0047] a second graph convolution unit configured to perform graph convolution operation on a pre-constructed user and device relationship graph through the second graph convolution network to obtain a second user vector corresponding to the first user and a first device vector corresponding to the first device, the user and device relationship graph being constructed according to historical association information between the plurality of users and a plurality of public face recognition devices, the plurality of public face recognition devices including the first device;

[0048] a third graph convolution unit configured to perform graph convolution operation on a pre-constructed device relationship graph through the third graph convolution network to obtain a second device vector corresponding to the first device, the device relationship graph being constructed based on geographical positions of the plurality of public face recognition devices;

[0049] a user and device vector determining unit configured to combine the first user vector and the second user vector to obtain a third user vector corresponding to the first user, and combine the first device vector and the second device vector to obtain a third device vector corresponding to the first device;

[0050] a network parameter updating unit configured to update network parameters of the first, second and third graph convolution networks so that a dot product value between the third user vector and the third device vector tends to be the first label value.

[0051] According to a third aspect, a computer readable storage medium is provided, having stored thereon a computer program which, when executed in a computer, causes the computer to perform the method according to the first aspect.

[0052] According to a fourth aspect, a computing device is provided, comprising a memory having stored therein executable code and a processor which, when executing the executable code, implements the method according to the first aspect.

[0053] With one or more of the methods, apparatuses, computing devices, and storage media in the above aspects, a risk judgment model for face recognition on a public recognition device can be trained, and based on the risk judgment model, the risk degree of face recognition on a public device can be quickly determined, and then the success rate of face recognition on a public recognition device can be improved according to the risk degree. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required to be used in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0055] Figure 1 A training scene schematic diagram of a risk judgment model for face recognition according to an embodiment of the present specification is shown;

[0056] Figure 2 A use scene schematic diagram of a risk judgment model for face recognition after training according to an embodiment of the present specification is shown;

[0057] Figure 3 A principle schematic diagram of a training method of a risk judgment model for face recognition according to an embodiment of the present specification is shown;

[0058] Figure 4 A flowchart of a training method of a risk judgment model for face recognition according to an embodiment of the present specification is shown;

[0059] Figure 5 A schematic diagram of a user similarity relationship graph according to an embodiment of the present specification is shown;

[0060] Figure 6 A schematic diagram of a user and device relationship graph according to an embodiment of the present specification is shown;

[0061] Figure 7 A schematic diagram of a device relationship graph according to an embodiment of the present specification is shown;

[0062] Figure 8 A schematic diagram showing a manner of obtaining a user representation vector according to an embodiment of the present specification is shown.

[0063] Figure 9 A structural diagram of a training device for a risk judgment model for face recognition according to an embodiment of the present specification is shown. DETAILED DESCRIPTION

[0064] The scheme provided in the present specification will be described below in conjunction with the accompanying drawings.

[0065] As described previously, the current equipment for face recognition is rapidly developing from personal terminal equipment mainly used for online face recognition scenarios to public recognition equipment mainly used for offline face recognition scenarios. However, there are the following problems in using face recognition technology in public recognition equipment: due to the fact that the number of users facing the public recognition equipment increases by a huge number of orders of magnitude compared with the personal terminal equipment, the difficulty of face recognition is increased, and the risk of face recognition on the public recognition equipment is significantly increased.

[0066] Specifically, in online facial recognition scenarios, users usually perform facial recognition verification through personal devices, and the number of accounts logged in / bound to the personal devices is usually small, such as the user himself / herself or a few other people closely related to the user. Therefore, there is usually a 1:1 relationship between the personal devices and the number of accounts logged in / bound to the personal devices. Further, when the user performs facial recognition in an online scenario, it is usually only necessary to verify the similarity between the collected face and the face of the bound account, and the recognition result can be obtained. Under this condition, the computer facial comparison algorithm used for recognition can only meet the performance required under the condition that the number of devices and users is approximately 1:1. Unlike this, in offline facial recognition scenarios, due to the public nature of the devices, there is not a 1:1 relationship between the devices and the accounts, but a 1:N relationship. That is, when the device performs facial recognition, it often needs to compare the collected face with the faces of all possible accounts to find the correct face and determine the account, and therefore there is also a 1:N relationship between the collected face and the faces of all possible accounts. N can represent the total number of possible accounts, or the total number of users in the user face library. In actual offline facial recognition scenarios, N can reach a very large order of magnitude. For example, in a completely open national offline facial recognition system, the order of magnitude of N can correspond to the total population, so N can reach hundreds of millions or even billions. In this case, as the user face library increases, the facial recognition error probability will increase significantly. The reason is that when N is large enough, when the face is swiped, there is a high probability that one or more faces similar to the face features of the face swiped can be found in the face library, resulting in misidentification. Generally, if a fixed computer vision-face image comparison algorithm / model is used, as the size N of the face library increases, the misidentification rate in the 1:N recognition scenario will increase in direct proportion to the increase in N. When N reaches a certain massive scale, the algorithm / model may no longer be able to meet the application.

[0067] Currently, there are only two solutions to the above problems: the first solution mainly improves the performance of the image comparison algorithm, i.e., improves the accuracy of 1:1 image comparison, to improve the accuracy of 1:N image comparison. However, there are still problems as follows: the misidentification rate of the image comparison algorithm is still approximately proportional to N, and when 1:N comparison is performed and N is large, the performance improvement of the comparison algorithm is usually difficult to offset the negative effects caused by the expansion of N. In addition, there is a bottleneck in the performance of the current image comparison algorithm, and it is difficult to directly greatly enhance the performance of the image comparison algorithm, i.e., it is difficult to improve the performance of 1:N comparison by improving the performance of 1:1 comparison algorithm. The second solution mainly obtains multiple users with face images similar to the collected face image in the process of 1:N image comparison, and then determines the correct user among these users through, for example, a judgment rule / machine learning model. For example, first perform image comparison, and then determine which user can pass safely through the judgment rule / machine learning model for the first and second users in the image comparison ranking. However, there are still problems as follows: when N is large, there is a high probability that the image comparison scores of the first and second users in the image comparison ranking are very close, and on this basis, the distinguishing effect of the rule-based judgment or model classification is often not satisfactory, and often causes a high business disturbance rate (interception rate), resulting in a poor experience of the face recognition user.

[0068] To solve the above technical problems, the embodiments of the present specification provide a training method of a risk judgment model for face recognition. The risk judgment model trained by the method can be used to judge the recognition error risk (or error probability) of face recognition performed on a public face recognition device. As shown in the Figure 1 The risk judgment model can be deployed in server A during training, for example, server A can be a physical server containing an independent host, or a virtual server, a cloud server, etc. carried by a host cluster. In the training stage, the risk judgment model can obtain samples for model training stored locally in server A, or obtain training samples from, for example, other storage servers. The samples for model training can include the recognition results (recognized users) of multiple face recognitions performed on multiple public face recognition devices, and labels indicating whether the recognition results are correct. Different training samples can be obtained according to historical face recognition records, and can also be obtained by statistical methods according to historical face recognition records. The embodiments of the present specification do not limit this. After the risk judgment model is trained, it can be used in, for example, Figure 2 the face recognition scene as shown in the Figure 2As shown, the trained risk judgment model can be deployed in server C for example, when an unknown user X performs face recognition through a public face recognition device - device 2 for example, the user can be identified as user 2 through a face recognition model deployed on server B for example, in order to determine the error probability of this identification, the identification of user 2 and the identification of device 2 can be sent to the risk judgment model deployed on server C, and the error probability of this identification can be determined through the risk judgment model. In different embodiments, the face recognition model and the trained risk judgment model can also be different and can be deployed in different ways. For example, in one embodiment, the face recognition model and the trained risk judgment model can also be deployed on the same server (for example, server C). In another embodiment, the face recognition model can also be deployed in a public face recognition device for example.

[0069] Figure 3 A schematic diagram showing the principle of a training method of a risk judgment model for face recognition according to an embodiment of the present specification is shown as follows. Figure 3 As shown, the model at least includes first, second and third graph convolutional neural networks. First, user similarity relationship graphs, user and device relationship graphs, and device relationship graphs can be constructed according to the similarity of the face images of a plurality of users, the historical association of these users on a plurality of public devices (for example, historical operation data of the users on the public devices), and the positional relationship between the public devices for example. Then, graph convolutional neural networks are used to perform graph convolutional operations on the user similarity relationship graphs, the user and device relationship graphs, and the device relationship graphs respectively. Specifically, as shown in Figure 3 For example, the first graph convolutional neural network is used to perform graph convolutional operations on the user similarity relationship graphs, to obtain first user vectors corresponding to each user respectively The second graph convolutional neural network is used to perform graph convolutional operations on the user and device relationship graphs, to obtain second user vectors corresponding to each user respectively and first device vectors corresponding to each device respectively The third graph convolutional neural network is used to perform graph convolutional operations on the device relationship graphs, to obtain second device vectors corresponding to each device respectively Then, the first and second user vectors of each user can be combined to obtain third user vectors h Ui corresponding to each user, and the first and second user vectors of each device can be combined to obtain third device vectors h Dj .

[0070] Then, several training samples can be obtained, and the first, second, and graph convolutional networks can be trained based on these samples. Specifically, any sample may include, for example, an identifier of a user (e.g., user Ui), an identifier of a device (e.g., device Uj), and a label value Lq indicating whether a recognition of user Ui on device Uj was correct / incorrect (e.g., a label value of 1 if the recognition was failed, and a label value of 0 if the recognition was successful). For any sample, the third user vector h corresponding to user Ui and device Uj obtained in the above steps can be used. Ui and the third device vector h Dj We can get h Ui and h Dj dot product value by The network parameters of the first, second, and graph convolutional networks are updated with the label value Lq as the target.

[0071] After the model is trained, representation vectors for each user (corresponding to the third user vector in the training process, which can also be called the user's misidentification risk vector since it is actually used to represent the misidentification risk at the user level) and representation vectors for each device (corresponding to the third device vector in the training process, which can also be called the device's misidentification risk vector since it is actually used to represent the misidentification risk at the device level) can be obtained. For any facial recognition on a public device, the risk level of that facial recognition can be obtained by taking the dot product of the user's representation vector and the public device's representation vector. Based on this risk level, a further judgment scheme for that facial recognition can be easily determined. For example, if the risk level of that facial recognition is below a certain threshold, meaning the risk of recognizing the user on that device is very low, then the recognition can be considered successful. If the risk level of that facial recognition is above a certain threshold, meaning the risk of recognizing the user on that device is high, then further verification can be performed, for example, by increasing the image comparison threshold or by using other judgment rules or models, to reduce the risk of misjudgment.

[0072] Using the method, the following advantages are obtained: 1. The inter-user similarity relationship in a 1:N face recognition scenario, the relationship between the user and the device, and the relationship between the devices are represented by multiple isomorphic (for example, inter-user similarity graph, device relationship graph) and heterogeneous graphs (for example, user-device relationship graph), and then the graph convolution algorithm is used to learn these graphs to obtain user-dimension and device-dimension false recognition risk vectors, which directly depict the false recognition risk of each user and each device. Then, the false recognition risk of a single user in face recognition on each device can be accurately measured, that is, the risk of a face recognition can be quickly judged. 2. The method for judging the risk of face recognition is not limited by the size of N. Therefore, when N is large, the method has obvious recognition performance advantages over the above-mentioned other schemes. 3. The method can improve the accuracy of false recognition risk judgment in face recognition, and the improvement of the accuracy of false recognition risk judgment can significantly reduce false interception, thereby effectively reducing the business disturbance rate (or interception rate) in face recognition and ensuring user experience.

[0073] The detailed process of the method is further described below. Figure 4 A flowchart of a training method of a risk judgment model for face recognition according to an embodiment of the present specification is shown. The risk judgment model at least includes first, second and third graph convolution networks, as shown in Figure 4 The method at least includes the following steps:

[0074] Step S401, a first sample is obtained, the first sample including an identification of a first user, an identification of a first device, and a first label value, the identification of the first user being a first recognition result output by the first device in face recognition, and the first label value being used to indicate whether the first recognition result is a false recognition result;

[0075] Step S403, the first graph convolution network is used to perform graph convolution operation on a pre-constructed user similarity graph to obtain a first user vector corresponding to the first user; the user similarity graph is constructed according to a first similarity between a plurality of users, and the first similarity is determined based on face images of the plurality of users, the plurality of users including the first user;

[0076] Step S405, the second graph convolution network is used to perform graph convolution operation on a pre-constructed user-device relationship graph to obtain a second user vector corresponding to the first user and a first device vector corresponding to the first device; the user-device relationship graph is constructed according to historical association information between the plurality of users and a plurality of public face recognition devices, and the plurality of public face recognition devices include the first device.

[0077] In step S407, the third graph convolutional network is used to perform graph convolutional operation on the pre-constructed device relationship graph to obtain a second device vector corresponding to the first device; the device relationship graph is constructed based on geographical positions of the plurality of public facial recognition devices.

[0078] In step S409, a third user vector corresponding to the first user is obtained by combining the first user vector and the second user vector; and a third device vector corresponding to the first device is obtained by combining the first device vector and the second device vector.

[0079] In step S411, the network parameters of the first, second and third graph convolutional networks are updated so that the dot product value between the third user vector and the third device vector tends to be the first label value.

[0080] First, in step S401, a first sample is obtained, the first sample including an identification of a first user, an identification of a first device, and a first label value, the identification of the first user being a first recognition result output by the first device performing facial recognition, and the first label value being used to indicate whether the first recognition result is an erroneous recognition result.

[0081] In this step, the obtained training sample can be composed of an identification of a user, an identification of a public facial recognition device, and a label value, the label value indicating whether a facial recognition performed by the public facial recognition device on the user is correct or erroneous. In different embodiments, the specific manner of obtaining the training sample can be different. In one embodiment, the training sample can be obtained from historical facial recognition data. For example, according to an erroneous recognition of user U1 on device D1 in the past, a training sample (e.g., a positive sample) can be obtained, which is composed of the identification of user U1, the identification of device D1, and the label value 1 (indicating erroneous recognition). According to a correct recognition of user U2 on device D2 in the past, another training negative sample (e.g., a negative sample) can be obtained, which is composed of the identification of user U2, the identification of device D2, and the label value 0 (indicating correct recognition). In one embodiment, the training sample can also be obtained by expanding the historical facial recognition data through a statistical method, which is not limited in the present specification. For example, in a specific example, the erroneous recognition of user U1 on device D1 in the past, and users U3 and U4 similar to U1 can be used to construct a plurality of training samples indicating the erroneous recognitions of U1, U3 and U4 on D1, respectively.

[0082] And, in step S403, a first user vector corresponding to the first user is obtained by performing a graph convolution operation on a pre-constructed user similarity graph by using a first graph convolution network; the user similarity graph is constructed according to a first similarity between a plurality of users, and the first similarity is determined based on facial images of the plurality of users, and the plurality of users include the first user.

[0083] In this step, the user similarity graph can be constructed according to pre-acquired facial images of a plurality of users. Specifically, in an embodiment, the facial images of the plurality of users can be compared to determine a facial similarity (first similarity) therebetween, and then a user similarity graph can be constructed according to the facial similarity. According to an implementation, the user similarity graph can include a plurality of user nodes and edges between the user nodes, the plurality of user nodes respectively correspond to the plurality of users, and the edges correspond to the first similarity between the users. As shown in the left graph (a) of FIG. 1. Figure 5

[0084] According to different implementations, the user similarity graph can also be constructed according to other user data or attributes other than facial data. For example, in an embodiment, the user similarity graph can be constructed according to a first similarity and a second similarity between a plurality of users pre-acquired, and the second similarity can be determined according to ages of the plurality of users and / or a number of times of using the same network, as shown in the right graph (b) of FIG. 1. Figure 5

[0085] In this step, by performing a graph convolution operation on the user similarity graph by using the first graph convolution network, a first user representation vector corresponding to each user in the user similarity graph can be obtained, and the first user representation vector includes a first user vector corresponding to the first user.

[0086] A graph convolution network (Graph Convolution Network) generally refers to a convolutional neural network capable of directly acting on a graph and learning structural information of the graph. Through the graph convolution network, a representation feature (such as a representation vector) of a node or a subgraph in the graph can be obtained, for example. In different embodiments, the first graph convolution network can be a graph convolution network of different specific types, and the present specification does not limit the specific type of the first graph convolution network.

[0087] And, in step S405, a second user vector corresponding to the first user and a first device vector corresponding to the first device are obtained by performing a graph convolution operation on a pre-constructed user-device relationship graph by using a second graph convolution network; the user-device relationship graph is constructed according to historical association information between a plurality of users and a plurality of public facial recognition devices, and the first device is included in the plurality of public facial recognition devices. ​​

[0088] In this step, the user-device relationship graph can be constructed in advance according to historical association information between a plurality of users and a plurality of public facial recognition devices. In different embodiments, the specific association between the users and the devices represented by the historical association information can be different. For example, in one embodiment, the historical association information can represent one or more of the face swiping / authentication / payment behaviors of the users through the public facial recognition devices, and in another embodiment, the historical association information can also be the coincidence relationship of the users and the public facial recognition devices in the spatial position at the historical time point.

[0089] According to an embodiment, the user-device relationship graph can include a plurality of user nodes, a plurality of device nodes, and edges between the user nodes and the device nodes, the plurality of user nodes respectively corresponding to a plurality of users, the plurality of device nodes respectively corresponding to the plurality of public facial recognition devices, and the edges corresponding to the historical association information between the users and the public facial recognition devices, as shown in Figure 6

[0090] In this step, the graph convolution operation is performed on the user-device relationship graph by the second graph convolution network, so as to obtain the second user representation vector corresponding to each user in the user-device relationship graph (including the second user vector corresponding to the first user), and the first device representation vector corresponding to each device (including the first device vector corresponding to the first device).

[0091] Similarly to the second graph convolution network, in different embodiments, the second graph convolution network can also be a graph convolution network of different specific types, and the specific type of the second graph convolution network is not limited in the present specification.

[0092] In step S407, the second device vector corresponding to the first device is obtained by performing the graph convolution operation on the pre-constructed device relationship graph by the third graph convolution network; and the device relationship graph is constructed based on the geographical positions of the plurality of public facial recognition devices.

[0093] In this step, the device relationship graph can be constructed in advance according to the positional relationship existing between a plurality of public facial recognition devices. In different embodiments, the specific positional relationship indicated by the first positional relationship can be different. In one specific embodiment, the first positional relationship is used to indicate that the distance between the public facial recognition devices is less than a first threshold. According to an embodiment, the device relationship graph can include a plurality of device nodes and edges between the plurality of device nodes, the plurality of device nodes respectively corresponding to the plurality of public facial recognition devices, and the edges corresponding to the first positional relationship existing between the public facial recognition devices, as shown in Figure 7 ​As shown in the middle left part (a). In a specific embodiment, the device relationship graph can also be constructed according to or in combination with other relationships between devices. In an example, the common user relationship between different devices can be determined according to, for example, the number of common users who have used different devices, and then the device relationship graph can be constructed in combination with the common user relationship, such as Figure 7 As shown in the middle right part (b).

[0094] In this step, the graph convolution operation is performed on the device relationship graph by the third graph convolution network, and the second device representation vector corresponding to each device in the device relationship graph (including the first device vector corresponding to the first device) can be obtained.

[0095] Similar to the first and second graph convolution networks, in different embodiments, the second graph convolution network can also be a graph convolution network of different specific types, and the specific type of the second graph convolution network is not limited in this specification.

[0096] It should be noted that in different embodiments, the above steps S403, S405 and S407 are not necessarily executed in sequence. For example, in an embodiment, steps S403, S405 and S407 can be executed in parallel. In an embodiment, steps S403, S405 and S407 can also be executed in any order. In addition, in different embodiments, the representation vectors of each user and device can be obtained first, and then the training samples can be obtained. That is, in some embodiments, step S401 is not necessarily executed before steps S403, S405 and S407.

[0097] After obtaining the first user vector and the second user vector, and the first device vector and the second device vector, in step S409, the third user vector corresponding to the first user can be obtained in combination with the first user vector and the second user vector; and the third device vector corresponding to the first device can be obtained in combination with the first device vector and the second device vector.

[0098] In different embodiments, the specific manner of combining the first user vector and the second user vector, and the specific manner of combining the first device vector and the second device vector, can be different. In an embodiment, the value of each dimension of the third user vector can be obtained according to the superimposed value / average value of the values of each dimension of the first user vector and the second user vector, such as Figure 8In one embodiment, the values of the first user vector and the second user vector in each dimension are added to obtain the values of the third user vector in each dimension. In one specific embodiment, the added values of the values of the first user vector and the second user vector in each dimension can be weighted added values of the values of the first user vector and the second user vector in each dimension. In different embodiments, the weighted weights in the weighted added values of the first user vector and the second user vector can be determined based on different specific manners. In another specific embodiment, the risk judgment model can further include a first attention network. The weighted weights of the values of the first user vector and the second user vector in each dimension in the weighted added values can be determined based on the first attention network, for example Figure 3 as shown in the above.

[0099] In one embodiment, the values of the third device vector in each dimension can also be obtained according to the added / averaged values of the values of the first device vector and the second device vector in each dimension. In one specific embodiment, the added values of the values of the first device vector and the second device vector in each dimension can be weighted added values of the values of the first device vector and the second device vector in each dimension. In different embodiments, the weighted weights in the weighted added values of the first device vector and the second user device can be determined based on different specific manners. In another specific embodiment, the risk judgment model can further include a second attention network, and the weighted weights of the values of the first device vector and the second device vector in each dimension in the weighted added values can be determined based on the second attention network, for example Figure 3 as shown in the above.

[0100] Thereafter, in step S411, the network parameters of the first, second and third graph convolution networks are updated with the purpose of making the dot product value between the third user vector and the third device vector tend to the first label value.

[0101] In this step, the risk judgment model is trained with the purpose of making the dot product value tend to the first label value, and at least the network parameters of the first, second and third graph convolution networks are updated. In different embodiments, different specific training manners can be used, which are not limited in the present specification. For example, in one embodiment, a risk loss can be determined according to the dot product value and the first label value; and the network parameters of the first, second and third graph convolution networks are updated with the purpose of making the risk loss tend to be smaller. In the above embodiments including the first and / or second attention network, the network parameters of the first, second, third graph convolution networks, and the first attention network and / or the second attention network can also be updated.

[0102] After the risk judgment model is trained, the risk degree of a face recognition on a public device can be quickly determined by using the risk judgment model. For example, if a second user is recognized in a face recognition on a second device, the misrecognition risk vector of the second user and the misrecognition risk vector of the second device can be determined based on the risk judgment model, and the risk degree (misrecognition probability) of the face recognition can be determined according to the misrecognition risk vectors of the second user and the second device. Therefore, in an embodiment, the plurality of users can further include the second user, and the plurality of public face recognition devices can further include the second device; a fourth user vector of the second user can be obtained by performing graph convolution operation on the user similarity relation graph by using the updated first graph convolution network; a fifth user vector of the second user can be obtained by performing graph convolution operation on the user and device relation graph by using the updated second graph convolution network; and a sixth user vector corresponding to the second user can be obtained by combining the fourth user vector and the fifth user vector. In another embodiment, a fourth device vector of the second device can be obtained by performing graph convolution operation on the user and device relation graph by using the updated second graph convolution network; a fifth device vector of the second device can be obtained by performing graph convolution operation on the device relation graph by using the updated third graph convolution network; and a sixth user vector corresponding to the second device can be obtained by combining the fourth device vector and the fifth device vector. In yet another embodiment, a second recognition result output by the face recognition performed by the second device can be obtained, the second recognition result including an identifier of the second user; a sixth user vector corresponding to the second user and a sixth device vector corresponding to the second device can be confirmed according to the identifier of the second user and the identifier of the second device respectively; and a probability that the second recognition result is a false recognition result can be determined according to the sixth user vector and the sixth device vector.

[0103] According to different embodiments, the specific manners of obtaining the misrecognition risk vector of the second user and the fourth device vector corresponding to the second device based on the risk judgment model can also be different. For example, according to an embodiment, the misrecognition risk vector of the second user and the misrecognition risk vector of the second device can be obtained according to the pre-constructed user similarity relation graph, the user and device relation graph, the device relation graph, and the trained risk judgment model before the second user is recognized on the second device. Further, the risk degree of the second recognition can be determined according to the dot product value of the misrecognition risk vector of the second user and the misrecognition risk vector of the second device. In addition, according to an embodiment, the misrecognition risk vector of the second user and the misrecognition risk vector of the second device, or the dot product value thereof, can be used as the recognition risk representation in the user and device dimensions, or a combination thereof, and can also be used as the input feature of other judgment models to further verify the recognition risk in other manners and further reduce the misjudgment risk.

[0104] According to another aspect, embodiments further provide a device for training a risk judgment model for face recognition. Figure 9 A structural diagram of a device for training a risk judgment model for face recognition according to an embodiment of the present specification is shown in FIG. 9. As shown in FIG. 9, the device 900 includes: Figure 9

[0105] A sample obtaining unit 91 is configured to obtain a first sample, the first sample including an identification of a first user, an identification of a first device, and a first label value, the identification of the first user being a first recognition result output by the first device performing face recognition, the first label value being used to indicate whether the first recognition result is an erroneous recognition result;

[0106] A first graph convolution unit 92 is configured to perform graph convolution operation on a pre-constructed user similarity relation graph by using the first graph convolution network to obtain a first user vector corresponding to the first user, the user similarity relation graph being constructed according to a first similarity between a plurality of users, the first similarity being determined based on face images of the plurality of users, the plurality of users including the first user;

[0107] A second graph convolution unit 93 is configured to perform graph convolution operation on a pre-constructed user and device relation graph by using the second graph convolution network to obtain a second user vector corresponding to the first user and a first device vector corresponding to the first device, the user and device relation graph being constructed according to historical association information between the plurality of users and a plurality of public face recognition devices, the plurality of public face recognition devices including the first device;

[0108] A third graph convolution unit 94 is configured to perform graph convolution operation on a pre-constructed device relation graph by using the third graph convolution network to obtain a second device vector corresponding to the first device, the device relation graph being constructed based on geographical positions of the plurality of public face recognition devices;

[0109] A user and device vector determining unit 95 is configured to combine the first user vector and the second user vector to obtain a third user vector corresponding to the first user, and combine the first device vector and the second device vector to obtain a third device vector corresponding to the first device;

[0110] A network parameter updating unit 96 is configured to update network parameters of the first, second, and third graph convolution networks in order to make a dot product value between the third user vector and the third device vector tend to the first label value.

[0111] ​In another aspect of the specification, a computer readable storage medium is provided, having stored thereon a computer program, which, when executed in a computer, causes the computer to perform any of the above methods.

[0112] In yet another aspect of the specification, a computing device is provided, comprising a memory having stored therein executable code and a processor that, when executing the executable code, implements any of the above methods.

[0113] It should be understood that the terms "first", "second" and the like in the description do not necessarily connote any temporal or chronological order, but merely distinguish different concepts from one another.

[0114] Those skilled in the art should be aware that, in one or more examples described above, the functions described in the specification can be implemented in hardware, software, firmware or any combination thereof. When implemented in software, the functions can be stored in a computer readable medium or transmitted as one or more instructions or codes on a computer readable medium.

[0115] The above detailed description of the specification has further detailed the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above description is merely a specific implementation of the present application and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made on the basis of the technical solutions of the present application shall be included in the protection scope of the present application.

Claims

1. A method for training a risk judgment model for face recognition, the risk judgment model comprising at least a first, a second and a third graph convolutional network, the method comprising: obtaining a first sample comprising an identity of a first user, an identity of a first device, and a first label value, the identity of the first user being a first recognition result output by the first device for face recognition, the first label value indicating whether the first recognition result is a false recognition result; performing graph convolutional operation on a pre-constructed user similarity graph by the first graph convolutional network to obtain a first user vector corresponding to the first user, the user similarity graph being constructed according to a first similarity between a plurality of users, the first similarity being determined based on face images of the plurality of users, the plurality of users including the first user; performing graph convolutional operation on a pre-constructed user and device relationship graph by the second graph convolutional network to obtain a second user vector corresponding to the first user and a first device vector corresponding to the first device, the user and device relationship graph being constructed according to historical association information between the plurality of users and a plurality of public face recognition devices, the plurality of public face recognition devices including the first device; performing graph convolutional operation on a pre-constructed device relationship graph by the third graph convolutional network to obtain a second device vector corresponding to the first device, the device relationship graph being constructed based on geographical positions of the plurality of public face recognition devices; combining the first user vector and the second user vector to obtain a third user vector corresponding to the first user, and combining the first device vector and the second device vector to obtain a third device vector corresponding to the first device; and updating network parameters of the first, second and third graph convolutional networks so that a dot product value between the third user vector and the third device vector tends to the first label value. The combining the first user vector and the second user vector to obtain the third user vector corresponding to the first user comprises: obtaining values of each dimension of the third user vector according to superimposed values / average values of values of each dimension of the first user vector and the second user vector. The superimposed values of values of each dimension of the first user vector and the second user vector comprise weighted superimposed values of values of each dimension of the first user vector and the second user vector. The risk judgment model further comprises a first attention network. The weighted weights of values of each dimension of the first user vector and the second user vector in the weighted superimposed values are determined based on the first attention network. The updating the network parameters of the first, second and third graph convolutional networks so that the dot product value between the third user vector and the third device vector tends to the first label value comprises: updating network parameters of the first, second, third graph convolutional networks and the first attention network so that the dot product value between the third user vector and the third device vector tends to the first label value. ​ 2. The method of claim 1, wherein, ​ ​ 3. The method of claim 2, wherein, ​ 4. The method of claim 3, wherein, ​ ​ ​ ​ 5. The method of claim 1, wherein, The combining the first device vector and the second device vector obtains a third device vector corresponding to the first device, and the third device vector comprises: The third device vector is obtained according to the superimposed value / average value of the values in each dimension of the first device vector and the second device vector.

6. The method of claim 5, wherein, The superimposed value of the values in each dimension of the first device vector and the second device vector comprises a weighted superimposed value of the values in each dimension of the first device vector and the second device vector.

7. The method of claim 6, wherein, The risk judgment model further comprises a second attention network. The weighted weights of the values in each dimension of the first device vector and the second device vector in the weighted superimposed value are determined based on the second attention network. The network parameters of the first, second and third graph convolutional networks are updated in order to make the dot product value between the third user vector and the third device vector tend to the first label value, and the network parameters of the first, second, third graph convolutional networks and the second attention network are updated in order to make the dot product value between the third user vector and the third device vector tend to the first label value. The user similarity relationship graph comprises a plurality of user nodes corresponding to the plurality of users and edges between the user nodes, and the edges correspond to first similarities between the users.

8. The method of claim 1, wherein, The user and device relationship graph comprises a plurality of user nodes corresponding to the plurality of users, a plurality of device nodes corresponding to the plurality of public facial recognition devices, and edges between the user nodes and the device nodes, and the edges correspond to historical association information between the users and the public facial recognition devices.

9. The method of claim 1, wherein, The device relationship graph comprises a plurality of device nodes corresponding to the plurality of public facial recognition devices and edges between the device nodes, and the edges correspond to first positional relationships between the public facial recognition devices.

10. The method of claim 1, wherein, The first positional relationship is used to indicate that the distances between the public facial recognition devices are less than a first threshold value.

11. The method of claim 10, wherein, The historical association information between the user and the public facial recognition device comprises one or more of face recognition / authentication / payment behaviors of the user through the public facial recognition device.

12. The method of claim 1, wherein, The user similarity relationship graph is constructed according to the first similarities between the plurality of users, and the user similarity relationship graph is constructed according to the first similarities and second similarities between the plurality of users, the second similarities are determined according to ages of the plurality of users and / or numbers of times of using the same network.

13. The method of claim 1, wherein, The plurality of users further comprises a second user, the plurality of public facial recognition devices further comprises a second device, and the method further comprises:

14. The method of claim 1, wherein, A fourth user vector of the second user is obtained by performing graph convolutional operation on the user similarity relationship graph by the updated first graph convolutional network; A fifth user vector of the second user is obtained by performing graph convolutional operation on the user and device relationship graph by the updated second graph convolutional network; The fourth user vector and the fifth user vector are combined to obtain a sixth user vector corresponding to the second user. ​ 15. The method of claim 14, further comprising: obtaining a fourth device vector of a second device by performing graph convolution operation on the user-device relationship graph by the updated second graph convolution network; obtaining a fifth device vector of the second device by performing graph convolution operation on the device relationship graph by the updated third graph convolution network; combining the fourth device vector and the fifth device vector to obtain a sixth user vector corresponding to the second device.

16. The method of claim 15, further comprising: obtaining a second recognition result output by the second device performing face recognition, the second recognition result comprising an identity of a second user; identifying a sixth user vector corresponding to the second user and a sixth device vector corresponding to the second device according to the identity of the second user and the identity of the second device respectively; determining a probability that the second recognition result is a false recognition result according to the sixth user vector and the sixth device vector.

17. A device for training a risk judgment model for face recognition, the risk judgment model comprising at least first, second and third graph convolution networks, the device comprising: a sample obtaining unit configured to obtain a first sample, the first sample comprising an identity of a first user, an identity of a first device, and a first label value, the identity of the first user being a first recognition result output by the first device performing face recognition, the first label value being used to indicate whether the first recognition result is a false recognition result; a first graph convolution unit configured to obtain a first user vector corresponding to the first user by performing graph convolution operation on a pre-constructed user similarity relationship graph by the first graph convolution network, the user similarity relationship graph being constructed according to first similarities between a plurality of users, the first similarities being determined based on face images of the plurality of users, the plurality of users including the first user; a second graph convolution unit configured to obtain a second user vector corresponding to the first user and a first device vector corresponding to the first device by performing graph convolution operation on a pre-constructed user-device relationship graph by the second graph convolution network, the user-device relationship graph being constructed according to historical association information between the plurality of users and a plurality of public face recognition devices, the plurality of public face recognition devices including the first device; a third graph convolution unit configured to obtain a second device vector corresponding to the first device by performing graph convolution operation on a pre-constructed device relationship graph by the third graph convolution network, the device relationship graph being constructed based on geographical positions of the plurality of public face recognition devices; a vector determining unit configured to obtain a third user vector corresponding to the first user by combining the first user vector and the second user vector, and obtain a third device vector corresponding to the first device by combining the first device vector and the second device vector; an updating unit configured to update network parameters of the first, second and third graph convolution networks so that a dot product value between the third user vector and the third device vector tends to be the first label value.

18. A computer readable storage medium having stored thereon a computer program, which, when executed in a computer, causes the computer to carry out the method of any one of claims 1-16.

19. A computing device comprising a memory and a processor, the memory having stored therein executable code, which, when executed by the processor, carries out the method of any one of claims 1-16.

Images